RIP Optane
And are we gonna go with, like, is this gonna be, like, a New Orleans funeral? Is this gonna be, like, a Viking funeral? Like, which cultural folkway are we picking for this funeral?
Speaker 2:It's been years.
Speaker 1:God. It is such a piece of literature. Richard Nixon's body Richard Nixon has gone home to hell, and his body is not meant to be burned in a dumpster.
Speaker 2:I mean, it was just
Speaker 1:like it it's it's so I'm not I I don't know which way we're going to this on offense. I I did go and I relistened to the, the hype episode.
Speaker 2:Mhmm. Which was Did we did we mention Optane?
Speaker 1:Did we mention Optane? I believe your exact line was that Octane blew a DUI on hype.
Speaker 2:Well, there there's no better audience for my jokes than myself. We're doomed by 6 months.
Speaker 1:No. It was great. So no. No. We definitely talked about opt in.
Speaker 1:We then also used that. And, unfortunately, I got, like I managed to get, like, trolled bypassed us. In that, we also talked about HPE is the machine and the memristor. And Cole Frederick, who did the notes, I've got no idea how, but he dug up that HPE machine ad that I was describing. Did you watch that thing?
Speaker 2:I did. And,
Speaker 1:it's Yes.
Speaker 2:I I did. I think I tweeted it out afterwards or
Speaker 1:something. So glad that he found that because I could be convinced that I dreamt that. It is so insane. It's bonkers. And HPE has clearly, like, engaged a PR firm and paid them good money to scrub this thing off the planet.
Speaker 1:I mean, the one we're linking to is 100% illegal. No question. I mean, it is a like, that thing is waiting. So, like, I would say, like, view it, but also don't because I'm worried that it's gonna get the DMCA takedown. It's amazing.
Speaker 1:That whole thing is just, it
Speaker 2:It's like it's like revenge advertisement or something.
Speaker 1:Oh, maybe that's what it is. Yeah. That's exactly Like like
Speaker 2:some cringey ad you made a long time ago that was a big mistake
Speaker 3:It's like
Speaker 4:left around.
Speaker 1:God. No. This person keeps posting it to social media. You know, I that there's just no amount of restraining order that'll prevent them doing it. It but that thing is is absolutely bonkers.
Speaker 1:And it does kinda dovetail in, I feel, to, like, the the saga of opt in. I'm kinda, like, wondering where exactly we wanna, kind of tack in here. Here. So one one way I thought about tacking in is, first of all, the because there was an article that was written that kind of prompted this in the register. Wasn't very good.
Speaker 1:It was very I mean, my tweet was to adopt it and write this. I drew that article. It was not great. Not the best.
Speaker 2:I did read it, and it felt Sorry.
Speaker 1:One second.
Speaker 2:It it I mean, it felt like, yeah, like, like, a weirdly apologetic, like, you know, it never, you know, it was never appreciated in its time sort of thing. I don't know. It it felt much too
Speaker 1:It felt it felt much too fair. And I also in particular, it kind of implied that software is to blame for Optane's demise.
Speaker 5:And it's like,
Speaker 1:woah. Woah. Woah. Woah. Woah.
Speaker 1:Woah.
Speaker 3:Woah. Woah.
Speaker 1:The offer is not to blame. And to the contrary, I think that, like, you and I have been around people and we ourselves have been these people predicting the non volatility of main memory for as long as I've been in the
Speaker 2:industry. Yes.
Speaker 1:So my question actually, Tom, for you is has that been true for your career as well in terms of people believing that main memory will be nonvolatile at some point?
Speaker 5:Well, keep in mind, I got started with core memory, which is nonvolatile. Touche.
Speaker 4:Yeah. Fair.
Speaker 5:Alright. And somehow somehow there was no software required
Speaker 4:to make it all the way.
Speaker 1:Right. But so has that been your experience as well that people were have been kinda because I also feel that, like and, actually so, Tom, this is actually interesting. It's a good a good segue in terms of the physics changes that you have seen in your career. I mean, obviously, core memory to DRAM is a big one, and and an important one, for a lot of reasons. I was trying to think back some of the physics changes that we have seen.
Speaker 1:Because certainly, I I mean, Adam and I, Adam especially, on the absolute forefront of the real the most profound physics change I feel we've seen, which is the introduction of Flash.
Speaker 5:Yeah. Flash is huge.
Speaker 2:Yeah. I think that's gotta be true, but now I've got in the back of my mind some of the EEs working with us at Oxide who who, you know, may beg to differ on, you know, like, IBM transitioning over to copper or something like that. But certainly, as I'm aware of, it feels like the most significant change. Yeah.
Speaker 1:I mean, I was gonna
Speaker 5:yeah. The other the the other long running tension is between fiber and copper where people people are always saying, oh, you're gonna have to use fiber to get to those speeds, and then the copper guys come out with the latest thing.
Speaker 1:Yeah. That's a good point. And and and certainly, fiber and copper have been kind of least rocking one another, and each has advantages over the other. But the conference certainly has economic advantages and and not, so yeah. That no.
Speaker 1:That's that's a good point. I was also thinking, like, SMP, Tom. The I mean, just special polypropylene, just in general, the hitting the memory wall and needing to add another compute element, and then ultimately which I do I mean, I I kinda do consider chiplets to be a a physics exchange where we and in part because actually there were people and many people at Intel who believed that chiplets were impossible, that that was the wrong approach. And that proved to be the right so I I don't know. I would I would come to that give partial credit to that one.
Speaker 2:And and Arnie has joined us. So so maybe he can he can weigh in too with with his, you know, more than 2ยข perhaps.
Speaker 6:Well, I I was gonna say that the on the on the sort of interconnect side, there's always the oh, the integrated photonics or Yeah.
Speaker 7:Yeah. Yeah. Right.
Speaker 6:You know, integrated light is just around the corner around
Speaker 1:the corner. Soon.
Speaker 4:Well And
Speaker 6:I don't wanna spoil it too much, but I've been told that it is gonna take yet another couple more more years even though I was really thinking that it was gonna happen, but no. Not so much.
Speaker 1:Well and, Ari, you and I are both hugely bullish on integrated photonics.
Speaker 3:I don't
Speaker 1:have we dragged you down this this rabbit hole, Adam, on integrated photonics? No. I mean, it oh, it is a delicious rabbit hole. I mean, it is very exciting where you are effectively taking light straight into the into the dye. Right?
Speaker 1:And eliminating a whole and then that allows you to go to much higher speeds, personally, much lower power. And there are a bunch of reasons why integrated photonics would be super silicon photonics would be super exciting.
Speaker 6:Yeah. A lot lower power, but also much better reliability.
Speaker 1:Much better reliability.
Speaker 7:Yeah. The the the the
Speaker 6:integrated photonics that Intel has been using for their, their optics that you can buy today, they claim a 1,000 times better reliability. So 3 orders of magnitude less likely to fail than existing optics. Now we would obviously have to put that to the test, but that's definitely where you need to be if you want to put photonics in a chip and solder it down to a a board and, you know, expect it not to Yeah.
Speaker 5:But the the photonics is is great, but the lasers are still kinda sucky for because of temperature problems.
Speaker 6:Yeah. But that's so that's their claim to to that they they solve that by integrating that all on a substrate. So it's it's all part of the silicon. It's So there's no there's no separate parts there.
Speaker 3:Right.
Speaker 5:Which is I like the the the guys doing it with the off board laser providing the light.
Speaker 3:Well, so I
Speaker 1:was also gonna say, like like, the purely optical computer, I feel, was something. Because I remember in the I wanna say, like, the early nineties, there was a research group that had some extremely promising results with a purely optical computer. It was gonna be and that that thing just completely evaporated. I've got no idea what happened to that. Another one that was a heartbreaker for me personally, and, Adam, I know you've heard me go on about this for, like, fucking 2 decades now, but carbon nanotube based memory.
Speaker 1:And this company, Nantero, that had a, like, super promising and it's, like, still around. The nanotube based memory, they had this great observation that instead of trying to grow these tubes in a particular orientation, they would let them grow kind of, like, all over the place, and then they would use photo lift to zap the ones that were out of alignment, which felt like a really good idea than using, like, VanderWaals forces to, like, bend these things. It just felt great, and they had, like, our our intention is to mention 3 orders of magnitude because that does seem to be, like, the number of orders of magnitude that you need to really get attention. It was, like, 3 orders of magnitude denser than DRAM and 2 orders of magnitude faster than DRAM and then nonvolatile. It's like, oh, man.
Speaker 1:This is exciting. But, no, it has not happened. It's
Speaker 5:Well, there's, there's also Everspin. It's got nonvolatile, but they've never been able to get good densities or cost.
Speaker 1:Everspin is he called
Speaker 5:it? Uh-huh.
Speaker 2:Oh, yeah. Everspin. Right? S p I n?
Speaker 5:Yeah. Everspin.
Speaker 1:Yeah. Yeah. Yeah. Yeah. And we yeah.
Speaker 1:What this is MRAM. Okay. Yeah. That's another one.
Speaker 2:Yeah. Well, after, after we started seeing NAND in lots of places in the enterprise, don't you feel like there was this sort of, like, renaissance of exotic memories that that were people looking for the next thing in the hierarchy. Like, like, all these technologies we were talking about sort of had their own sort of moments of discussion, you know, in the sort of 2010 era, 2010 to 2015 era, as we were looking for the next step in the memory hierarchy.
Speaker 1:Yeah. I mean, I do think that, like, certainly, the, NAND actually really bursting into the memory hierarchy and becoming meaningful for basically all compute, from personal computing, obviously, handheld computing, and and server side computing. It did, I think, give a lot of inspiration to a bunch of these things or maybe a funding, maybe.
Speaker 2:You're right.
Speaker 1:And and the other one Tom, do you remember how graphic storage? This is another one of these where it's like, am I tripping? Did this actually happen? I swear IBM was talking about holographic storage.
Speaker 4:Oh.
Speaker 5:But Yeah. That definitely reached out.
Speaker 1:Oh, got it.
Speaker 5:I There's some 3 three d types optical thing. I don't know what that was. Holographic.
Speaker 1:Adam, I really thought you were just gonna fuck with me and be like, oh, was a walrus there, and were you on a boat? I mean, I I just thought you pushed them out because I as I was describing, I'm like, this actually sounds more dreamlike. But I no. No. They're alright.
Speaker 1:No. They're holographic data stories.
Speaker 6:Well, Brian, in that case, we both had the same dream because I also remember reading about that a long time ago.
Speaker 2:So And it's like, what We might be sharing
Speaker 1:You're right. Sharing the hallucination. And I do feel that, like, we, as technologists, it's incumbent upon us to remain optimistic about I mean, it's like, we don't wanna be too cynical about these about new physics because those things can be really relevant. And so I think I just wanna, like, set the stage that we, I would like to say collectively, are not anti alternative physics. Optane, do you see where we're going with this?
Speaker 1:I know. I mean, I just I I I just because I feel like that was kind of the tone of that article was that software kinda refused to adopt this thing. And I I really feel pretty emphatically that that's not the case.
Speaker 2:No. I think we wanted to believe. It just let
Speaker 3:us down.
Speaker 2:Yes. Jeffrey, you joined. Did you wanna hop in?
Speaker 6:I just wanted to say that, Microsoft Research is still working on holographic storage. There you go.
Speaker 1:Oh, that's good. Yeah. I mean well and and honestly, like, all these efforts are, I think, great and important. I think that we now all of that said, DRAM is really, really hard to beat. And if you're going up against DRAM, and it's also a moving target.
Speaker 1:It's getting better and better and better. It's been I mean, it's not just riding Moore's law, but it is, there are there's so much economic incentive to continue DRAM densities, and of course, it's very high speed.
Speaker 2:And and that's the key. I mean, the economic incentives are the key of what made, NAND pervasive. Right? Because prior to its ubiquity and level, like, I mean, you know, NAND had existed since the seventies or whatever, even SSDs. So when we started building the, ZFS storage appliance or what would later become the ZFS storage plants, we were kicking the tires on SSDs.
Speaker 2:And the first batch we got in, like, 2006, 2007, were all intended for ruggedized uses. Right. You know, show us, the form factor of your helicopter and we'll tell you the right SSD because, you know, obviously, you have a VIBE problem. And then it was the next batch where we started seeing high performance and high capacities, but it was all, you know, enabled by this economic shift.
Speaker 1:Oh, wait. Not all high performance. When was mister Shipbag in there? Do you remember mister Shipback? I
Speaker 2:I don't. I I do remember a a a drive affectionately noticed mister Shipback. What was its pathology? I mean, it would simply, like, go out to lunch for, like, a very long time.
Speaker 3:Mister Shipback had worse performance than my IBM PC XT.
Speaker 1:Was the I actually Seagate drive spec on that drive. It had 93
Speaker 4:millisecond rights.
Speaker 1:And yeah. Exactly. Well, because when when
Speaker 5:I I I I I played with the early batch of SSDs where you could get us so backed up that it could take 5 or 6 seconds for a ride.
Speaker 1:Jeez. Yeah. That's bad. Once I think at the end, Tom, there's kind of an interesting thing there about, like, the nature of the physics with the NAND. Like, the way that it works has certain implications for software.
Speaker 1:And I think it was actually really important to know how DAN works, know how it's operate. Like, know why it's called Flash. Right? I mean, the name Flash comes from the way effective data physics works. It reminded the inventor of of a Flash going off and effectively resetting a large bank.
Speaker 1:I I think I remember that correctly, right now?
Speaker 2:Yeah. Yeah. That's right. Yeah.
Speaker 1:And that was at Toshiba, I think, back in the day. Right? What what was what whatever Flash was. I think so.
Speaker 5:Yeah. Toshiba. Yeah.
Speaker 1:So sorry. So this leads us to I I I do wanna kinda get to the main event in terms of, Optane, nay, 3 d Crosspoint, nay, Apache Pass.
Speaker 2:Oh, that I see. I I couldn't remember the other code name. Because I remember coming to you at one point, probably in 2015 at Joyant, maybe 2016, and talking about, like 3 d cross point or something. And you being like, oh, do you mean Apache pass? Because you you had been read in by the by, like, folks making it, presumably.
Speaker 1:Read in is really strong.
Speaker 7:So Yeah.
Speaker 1:I mean, so we had I mean and I still, like, absolutely love our our intel rep. It's terrific. She's great. And one, to give us as early visibility as as she possibly could on things. And so she had given us early visibility to Apache PaaS.
Speaker 1:It'd be interesting to
Speaker 2:know for folks that were
Speaker 1:at, like, hyperscalers that may have had even that would have much more intimate relationships with Intel, or really just anybody that found out about Spectre meltdown, not on Hacker News on January 2, 2018. The, but, she was describing Apache Pass, and it's, like, intriguing in my first question. And it may have been actually, I think, many of our first questions was, that's great. How does it work? Like, what is it?
Speaker 1:Is it holographic storage? Is it face edge memory? Is it is it MRAM? Is it you know, what or is it something totally Intel was they were not gonna describe in fact, have still really not fully described what it is or how it works. And this, I think, we get to very serious failing, like, debilitating critical fatal failure number 1, is that Intel's refusal and its secrecy in terms of describing how it worked, I think, really limited what people could go do with it.
Speaker 1:Because, I mean, I I actually think that it's really important to understand the physics of this stuff so you can understand I mean, like, obviously, like, you know, I'm a software engineer. I don't actually like, I'm not gonna weigh in on this. I know. You know? I but I I I I don't wanna understand how it works so I can replicate it.
Speaker 1:I wanna understand how it works so I can understand what the software implications are and how to treat it and how to think about, you know, as we're kinda thinking about next gen stuff. And I they I we they would never answer how it worked. I don't know, Adam. Did you ever did anyone ever get an answer from from Intel?
Speaker 2:No. No. Never never an inkling how it worked. And but I but I think they did this weird position where they really thought it was gonna be the golden goose or the goose that go laid the golden neck. Right?
Speaker 2:They thought that this was going to be both ubiquitous and proprietary, which, like, that those things don't happen at the same time.
Speaker 1:Okay. That's And because hold on. That's a really important point, actually. Just like I think that that's actually pretty profound, what you just said, in terms of, like, ubiquitous and proprietary don't exist at the same time. Or at least not for long.
Speaker 1:No. I agree with you.
Speaker 2:It might be held, held attention momentarily, but they, but they really believe they did. And they were, because I correct me if I'm wrong, they were building special support for this into the chips they were making, like into the CPUs they were making.
Speaker 3:Yeah.
Speaker 2:And if you want it to play at the Optane, you know, super fast, you know, playground, it was gonna be Intel only when it came to CPUs.
Speaker 1:Yes. And
Speaker 5:and so so Facebook had a box,
Speaker 6:at least one, because I've I've I've never seen it because it was in a data center, but we did some experiments level 2 structures of an of an LSM
Speaker 5:to this.
Speaker 6:And level 2 structures of an of an LSM to this, but it never I don't think it ever materialized into anything more than a prototype, and I don't know why.
Speaker 1:But
Speaker 6:I know there was actual machine from Intel with with indeed a a Xeon CPU that was able to support this and that ran this. This was before we even knew what it was called.
Speaker 5:Yeah. Yeah. They they definitely had the platforms prototyped and available for developers and all that and, you know, became a real product, but I'm not sure it ever had a viable roadmap.
Speaker 1:And, Tom, did you use it or or, Ari, maybe when you at Facebook were using it, was this in which kind of embodiment was this? Because I think the the very confusion that Optane had about where do we belong is actually you can see that in the form factor where it was, like, it was made available in both a dim and an SSD form factor. But I think that the that that the initial vision certainly as it was, I believe, it was portrayed to us in whatever it was, 2014, 2015, was this is going to be a a a DIM form factor. This is gonna replace DIMMs.
Speaker 6:Yeah. We were we were supposed to look at it as a DIMM because what we're they were trying to do is, right, basically, just use the in memory structure, like like, the the in memory representation of the first levels of the of the LSM, you just didn't have to serialize them to disk. You could just leave them in memory, and they would always be in memory.
Speaker 1:Right.
Speaker 6:And so it would be and it would be crash safe and whatnot. And, like, there there were some claims made that were definitely pretty wild, but the that was the intent to use that as the as the the in memory representation and just never had to serialize these and therefore, you would avoid all those the overhead.
Speaker 5:Oh, yeah. And then sec so so secondly, though, the position was that PCI Express was gonna be too slow, and therefore, it had to be on the memory bus
Speaker 3:Right.
Speaker 5:You know, to to really show off what it could do.
Speaker 2:Arrian, this reminds me of, you you're talking about Facebook and and it's considering Optane. I had a very interesting conversation with Jeff Rothschild. This was probably in 2016 because I was, briefly an entrepreneur in residence and he was doing some diligence on a company and I was called in some due diligence on the same company. And, I found myself in the uncomfortable position of making a claim that, that Jeff really didn't care for. And the claim that I made was that for memory, persistence is sort of a parlor trick.
Speaker 2:And the reason I said that was that if you've got data that has persisted, that's important, and and it's only on one node. So how how useful is that in an enterprise setting? And he he corrected me saying that, like, the startup time for these nodes in your, in the Facebook infrastructure was critically important. Warming up those caches was was critically important. Like, I sort of agree with him, but it also feels like, and Brian, you've said this in the past, the volatility volatility of memory has its benefits in that you don't keep around some corrupted state persistently forever.
Speaker 1:I do think that memory being nonvolatile is both there there's terrific potential there, but also terrible peril. And because, I mean, if you've ever had I mean, and and, obviously, we have had in our careers, but in in many ways, like, the gnarliest bug I've ever been involved in debugging was a data corruption a kernel data corruption bug that managed to leap the fire line into CFS. We had we had a couple of instances where that wild kernel data corruption had instances where that wild kernel data corruption had corrupted a buffer that was on its way out the disk, and it took us so long to put those 2 and 2 together that, like, wait a minute. This I mean, it's I think it's actually one of the very few cases of I think it's actually one of only 2 cases of on disk CFS corruption I've had. And it was due to a and so the idea that an arbitrary memory corruption issue, now
Speaker 3:all of
Speaker 1:those become persistent. That is a scary idea for sure. And I'm like, I wanna think carefully before opening Pandora's box on that one. I mean, obviously, as I Ari, as you're talking about, as Jeff was talking about, there's great potential there too. I just don't know I don't think we would wanna make it transparently nonvolatile.
Speaker 1:I think that that would be perilous.
Speaker 6:I think that's part of why that whole experiment never really went much further or or definitely never went into production. Be because exactly that. If you want to protect against that potential corruption, then you're gonna have to in addition to what you're already doing in memory, you're gonna have to add more over on top of that. And now you're starting to chip away at the benefit. So you're gonna have to calculate some amount of parity and keep it on the side to potentially prepare your primary data structure.
Speaker 6:And now you're take chipping away at all those supposed benefits that you're gonna have, and you might as well have serialized it somewhere and just kept it.
Speaker 1:Yeah. I mean, are you not at the bottom. I feel like you're that's that's kind of another embodiment of what Adam is saying too about, like, you've got these all these other factors now that you're gonna have to consider because this is your single store effectively. And I like your point about yeah. Like, we we don't actually use, in software, this happens in hardware, but we don't actually, keep auxiliary data structures to allow us to repair our state in memory.
Speaker 1:But we're gonna need we would need to do that in a a world that's all nonvolatile. So, in are you do you know what the because I I gotta believe though also that the Adam, because you had another point that went out too in terms of it being Intel only. But but before you get there, I think the other problem is, like, just from a performance perspective, it just didn't
Speaker 2:deliver. Well, in a in a price performance perspective in particular. Right? It would I mean, the the beauty of NAND flash, you know, after iPhone and iPods and everything dropped the the the price of NAND so dramatically or or I credit for dropping the price of NAND, so dramatically was that it was like, it really sat nicely in the middle of that hierarchy between in terms of price performance between disk and DRAM. And I always felt like Octane was like a little bit faster than NAND and a little bit cheaper than DRAM, as opposed to
Speaker 5:And and and it was far overhyped for both of those roles.
Speaker 1:Yeah. And there's also
Speaker 6:Yeah. You couldn't get the density out of it to really use it for large so we like, looking at INGIS, it's like, okay. Well, I don't know how much data that is gonna be, potentially gigabytes of data, but because but you guys, you can go very deep in terms of levels there. So you you quickly either run out of space there or and then you you need to spill over the flash anyway, and then you, you know, the programming model doesn't work very well. So there's always all all these auxiliary things.
Speaker 6:Like, there wasn't really a position where it would naturally fit. And I think that's And they,
Speaker 5:they they only got to 5 12 gig DIMMs, but they required a normal DRAM DIMM paired with it. So the overall density was only twice as good as you could get with a 128
Speaker 1:Yeah. That's a good point, Tom. That that and that was just kinda burden of the form factor. So there are a bunch of the we and then you kinda get into, like, Optimus is, you know, this very as Apache passed, it's kinda very promising birth. But then it becomes delayed.
Speaker 1:I mean, it definitely took longer than they thought, which fine. Understandable. The economics didn't seem to deliver. The part of the reason that I kept asking how does it work, and I kept trying to explain it like, look. If you can't explain how it works, I'm not gonna adopt it because I can't just take a magic box.
Speaker 1:I I need to understand how it works. And in particular, the reason I wanted to understand how it works is that the question that I had was, does it wear out? And the answer, even though Intel has been very reluctant to say it, is yes. It does wear out. And it is my understanding.
Speaker 2:Really? Because I I never got experience. Like, I I
Speaker 5:Yeah. Well well, part part part of their PR blitz when it first came out was that it would not ever wear out. And then they had a major design change to to support wear leveling on on the DIMM stuff as it did wear out.
Speaker 1:Right. And so we
Speaker 6:Which which makes the whole discussion about corruption of your persistence.
Speaker 4:All the all the more nerve wracking.
Speaker 1:Right. It's like, why are you worried about it? It's not volatile. It's like, does it wear out? No.
Speaker 1:Well, maybe. It's like, wait a minute. What?
Speaker 2:No. It doesn't wear out. Just just be careful with how you write the stuff. Don't write
Speaker 4:to the same spot. Well
Speaker 2:So what would you say?
Speaker 3:Well, of
Speaker 5:course, there there's RowHammer, so you have to worry about Deepgram too.
Speaker 1:To to fair, when actually so I'm actually, Roeham is kind of an interesting point, Tom, where you've got something that is a technology that's very well understood, and still we are learning things and getting surprised by the way that it behaves. In that case, I mean, I don't because row number 1 actually, I don't think will destroy the rabbit hole, but it will allow you to read cells that you should be able to read. So the, but, yeah, that that that's a very good point. I I think that also at some point, they're like, no. No.
Speaker 1:It doesn't wear out. And then it's like, no. It has, like it's a 1000 times better than NAND. And you're like, wait a minute. Is it, like, would it be infinitely better than NAND if it doesn't wear out?
Speaker 1:Why is it only a 1000 times better
Speaker 4:than NAND?
Speaker 2:So it's like, you know, it's also a thousand times better than NAND? Like like, older NAND. Like, NAND that isn't trying to get That's right. You know, like, such such tiny photography.
Speaker 1:And then I just feel like there was a bunch of, like, kind of stuff that's, like, kinda slimy where it was like, I don't know. Like, it won't wear out for the the the depreciation time of the asset. And you're like, woah. Woah. Woah.
Speaker 1:Woah. Woah. Woah. Woah. Woah.
Speaker 1:Woah.
Speaker 3:Woah. Woah. Woah. Woah. Woah.
Speaker 3:Woah. Woah. Woah. Woah. Woah.
Speaker 3:Woah. Woah. Woah. Woah. Woah.
Speaker 3:Woah. Woah. Woah. Woah.
Speaker 1:Woah. That is not the question.
Speaker 2:That's a big ask
Speaker 1:That's a big asterisk. That's like a day like, sorry. When I asked would it wear out, like, that's not an accounting question. I mean
Speaker 2:No. It's like, no. The warranty is only 2 years. Right. It's like, no.
Speaker 2:Wait.
Speaker 3:That's why
Speaker 2:why do you mention that? Why are you talking about it?
Speaker 1:Because it won't wear out during the warranty period. Like, what do you not explain? What do you not understand about this, Greg? And then you're like, well, well, can you just explain how it works? I don't know.
Speaker 1:We can't explain how it works. But it's and it's also, like, okay. Is it phase change memory? No. It is not phase change memory.
Speaker 1:You're like, okay. Which Intel did I mean, Intel has and I think that the there was, I think the e times actually like, someone took one of these things apart and and validated that it actually is PCM. It is phase change memory, and it is using Gaussian glass, but it is not but it is there it is more complicated than that. But it ultimately it is using it is phase change memory. And I don't understand why Intel I don't know if they were concerned that if they just called it phase change memory, people would begin to ask a bunch of very legitimate follow-up questions.
Speaker 1:I I just I don't get it. And I think it was a huge so I think that was, like, kinda mistake number 1, honestly, is not talking candid. I I wish they talked candidly about it, left it as a as a very as a research project that they're putting a lot of money into as opposed to trying to really put it on a product road map.
Speaker 2:Well, and and
Speaker 5:Maybe they maybe they were afraid of attracting patent trolls if they named the technology.
Speaker 4:Could be.
Speaker 2:Oh, man. Someone's got a lot of lawyers. Yeah.
Speaker 8:I I And do
Speaker 2:you remember remember in terms of the product roadmap too, do you remember it it they they hyped its use as nonvolatile memory, but the first products were all extremely expensive, slightly faster SSDs.
Speaker 1:Right.
Speaker 2:And and then tried to push a narrative around this hybrid model. Yes. And where where we had been using NV backed DRAM in this kind of hybrid model. And it wasn't clear to me that, like, you needed that much, you know, that much log space intent log effectively.
Speaker 5:Well, plus all the new all the new SSDs, Deepgram caches anyway. So
Speaker 1:Totally. Well and we had I mean, that that was actually the the the SSD that that we affectionately named Logzilla, sold to us by the person who didn't go to jail, unlike his brother, I think. Right?
Speaker 2:That's right. That's right. Well, I mean, they're all part of of Aztec together. So I I But, yes, only one of them went to jail. Yes.
Speaker 1:What's the, the movie about the arms dealers? What is that? That is, I you know, I watched that not even that long ago.
Speaker 2:Lord of War.
Speaker 1:Lord of War. Don't you feel like we were kind of, like, playing out Lord of War with Flash Dives?
Speaker 2:No idea. Absolutely. Wait.
Speaker 7:Lord of
Speaker 4:War or Dogs?
Speaker 1:War Dogs.
Speaker 4:Was it
Speaker 2:funny or was it serious? No.
Speaker 1:No. It was War Dogs. It was like Okay. Yeah. I'm thinking War Dogs.
Speaker 1:Maybe you need to go watch Lord of War. No. It's it's it's War Dogs, which is which is listed, by the way, Ariane, as a war slash comedy. So there you go. Do you
Speaker 6:It's also a little it's dumb. So, yeah, it is It's
Speaker 1:comedy. But the the so maybe I should go watch Lord of War, but this is about out of War Dogs. Have you seen War Dogs? This is Jonah Hill.
Speaker 2:No. No. No. No.
Speaker 3:This is
Speaker 1:about these guys
Speaker 6:They're both great.
Speaker 1:That that that become basically arms dealers and get way over their skis. They kinda become the accidental arms dealers and kinda get into this very shady underworld. And I feel like with we were so early in Flash that it, I mean, as it turns out, the the company that was the leading company in this, STech, was involved in some really serious criminal activity in that they were stuffing the channel and lying about their results. And
Speaker 2:Lying about their results to The Street to to be clear. So public company.
Speaker 1:And, and did but, like, the I feel like of the Machete brothers, like, Mark did not go to I know I I think they, like, I think they were both, like, off on technicalities or something, but they clearly did it.
Speaker 2:Yeah. And then and then they had to, like, kinda do some arrangement where, Mark's brother, I can't remember the other one, had like, wasn't allowed to be the CEO of the company. Anyway, there was a bunch of there's a bunch of stuff.
Speaker 1:Elizabeth Holmes treatment. Yeah. So it but the but so that drive sorry. I didn't get us didn't mean to take us down. They the War Dogs played, but the Flash Dogs, that drive had the I mean, that would had DRAM on it.
Speaker 1:That was the whole point of that thing is that it was using DRAM to front front flash.
Speaker 2:Yeah. And this was the, this was actually techno so STEC was originally or, I guess, you know, when we stumbled into them at around like 2,006 was selling a lot of like consumer SSDs and consumer DRAM. And, you know, if you went to fries, it'd be packaged as tech and they acquired this company, GNU Tech from England, that was building these pretty, pretty neat SSDs with FPGAs on them, for for the, you know, for the FTL and for the for the controller. And then, you know, acquired that company and divested all of the less interesting parts of STech to really focus on this technology.
Speaker 1:And that particular device from our perspective didn't have crosshairs on necessarily DIMMs or even storage. It was actually battery backed DRAM that we were seeking to replace with that thing.
Speaker 2:Yeah. For for us, I mean, it was a it was a lifesaver in terms of having battery backed DRAM effectively that was then dual attached. So getting out of the yes. Sure. It's on the volatile, but how many machines can attach to it.
Speaker 2:But, for their channel stuffing partner, EMC, it was replacing short stroked 15 k RPM disks. Right. And I don't I don't even remember this. Do you want we priced it sort of reasonably, like, maybe twice what it cost us or something like that. Do you remember how EMC priced it?
Speaker 2:No. Same drive? What they did is they calculated the number of IOs in a 15 k RPN disk and then charged proportionally for these
Speaker 5:SSDs based
Speaker 3:on It's
Speaker 5:good to
Speaker 1:be kidding. Tactics.
Speaker 2:Absolutely. So these drives that we were selling for, like, I don't know, like, $1,000 were, like, no joke, like, $32,000 from EMC.
Speaker 1:I you know, god bless us tech for stuffing that channel. Who would stuff that channel? That channel deserves to be stuff.
Speaker 2:Yeah. Well, in in just just to to fill in the details there, it was like the, EMC bought a ton of them thinking they were gonna sell a ton of them. It turned out people didn't want them so much at $30,000 or what or whatever. And they said to s tech, you know, we have enough. Like, this is the last order.
Speaker 2:We're gonna make this order because you're twisting our arm, but this is the last order you're gonna get from us in a very long time. And then, and and then Aztec turned around and said to the street, you know, up into the right, everything is looking great. No obstacles on the horizon.
Speaker 1:Don't know what you're talking about. Yeah.
Speaker 2:And then and then and then the 2 brothers sold a a shit ton.
Speaker 1:Oh, yeah. Right. Then there's that. Then there's the wild profiteering from the inside transformation that the company is about
Speaker 5:to collapse.
Speaker 1:Yeah. Yeah. There you go. But not convicted, I think, if I recall correctly.
Speaker 3:I think
Speaker 1:that they were Yeah. But I definitely remember, like, a bunch of false journal exposes, but, like, hey. Wait a minute. Wasn't that the guy that was in our office? I gotta get to that guy.
Speaker 2:I mean, our current partner. Yes.
Speaker 1:So listen. I'm just trying to say that, like, when you're talking about new physics, it it's gonna like, it's wild, wild west out there. It is the I'm not sure what the the the the obtained equivalent of this is. But the I also feel that, like, they they they were confused about who they were competing with. Is this better DRAM?
Speaker 1:Is it a faster SSD? And then there was the Tom, did you have where they were making Optane available in these DIMM form factors that had both DRAM and Optane on them. And their firmware on the DIMM was going to take care of of caching the the the it it it was going to basically syncing things out from the DIMM. So they were fronting the they were doing what the s tech drive
Speaker 5:was doing. Right.
Speaker 1:I'm like, I do not want that kind of firmware on my DIMM.
Speaker 2:That is terrifying. It it's like these, you know, like these, hard drives that have a little bit of NAND in them. I I think they've kinda fallen out of fashion.
Speaker 5:But you know the same thing that that that that was just the prototype for the the real Apache Pass product where the firmware just moved into the CPU chipset. Right. And it's still still dark matter in there that you
Speaker 1:can't So the okay. So, Tom, this gets us to, I think and I I muted on this earlier. But this is another fatal failing of this is that it required and the the Tom, was that software that logic, was that in the management engine? I I feel like Robert knows the answer to this. So, Robert, if you're listening to something.
Speaker 5:No. I think it's a whole a whole new subsystem in the chipsets that support that.
Speaker 1:And in the memory controller effectively. That would yeah. So the the the upshot of this was that you could not use Optane on anything that was not from Intel. And that was really debilitating because now you're a lot if you get locked into Optane, you're also locked into Intel. And for us at Oxide, we actually did take a look at update.
Speaker 1:I I don't know. I feel like we're taking a look at update when you right? When you joined, we were looking at update.
Speaker 6:Brian, you can also look at this the other way around. If you buy something else on Intel, you don't accidentally buy Optane.
Speaker 4:What do
Speaker 1:you mean? If it
Speaker 6:Like, oh, I don't want Optane. To make sure, I'll just buy AMD
Speaker 4:because I
Speaker 1:can never Right. Oh, yeah. Exactly. Right. Exactly.
Speaker 1:You you you can't actually stuff update into my part. Yeah. Yeah. Fair enough. Yes.
Speaker 5:This won't just won't Yeah. Yeah.
Speaker 1:Fair enough. But, Adam, we were looking at update. Right? I feel. We did look at it.
Speaker 6:Yeah. Very briefly.
Speaker 2:You know what? I choose to not remember that. Like, I I I don't remember that, but I think it's probably by chance. That It was a it was
Speaker 6:a one one of the bullet points on the list of, like, in the shootout between the CPUs, we held the door open for a little bit. I'm like, oh, this could be possible, would be possible if we had Intel CPUs.
Speaker 1:But I feel
Speaker 2:And then we
Speaker 6:concluded now. We don't want it anyway.
Speaker 1:We don't want it anyway. Yeah. Yeah. Yeah. I kinda feel like the and I think that, like, if Optune had lived up I mean, it would have been a different world, and they might have been able to pull off some of this stuff had it been just absolutely just DRAM beating performance density and so on and economics.
Speaker 1:Like, that would be maybe you could pull this off, but definitely not with it being a laggard. And and for us, it was just like it was not now there have been companies that have bet completely on Optane. Did you see did you see this item? Do you VAST data? Do you remember these these guys?
Speaker 1:They're the right.
Speaker 2:Yes. Yes. No. I I I do know about VAST because, our our friend of Oxide, who who may join us had been doing some diligence on them. And, and and because they were doing doing, Octane as, like, a caching tier or whatever.
Speaker 2:Not as it's like a hybrid storage, not the whole thing.
Speaker 5:Yes. Right.
Speaker 1:And VAST made a huge deal about the fact they're using an update. So when I saw the the kind of
Speaker 3:the announcement, whatever it was
Speaker 1:last week or whatever, I'm like, someone needs to do a welfare check on VAST to make sure, like, how's VAST? Do you see their blog entry is like, things couldn't be better here at VAST, and I don't know what everyone is talking about.
Speaker 2:Good. No. No. No.
Speaker 1:Yeah. No. No. No. I actually like you know what?
Speaker 1:Hey. I admire it. I'm like, they were you know, and the, you know, they they the title of the blog entry is Fast Position on Intel Optane Discontinuation, colon, nothing changes. It's like, I mean, okay. Something's changed.
Speaker 1:It's okay. Like, you I think, like, Colin, we're gonna live would be fine.
Speaker 2:Everything is fine. Don't worry.
Speaker 1:It's like but, like and I clearly, like, I'm very sympathetic with any fellow startup, especially one that's taking a big swing. And so, like, obviously, you know, I want them to live even though they
Speaker 3:they
Speaker 1:bet it all on Optane. But, again, I guess I guess they didn't bet it all on Optane. I guess that's the point. It's like they actually they've got they do have some of their chips, and they're gonna win it back. But actually
Speaker 5:Well, they Yeah. I mean, their their principle
Speaker 6:Oh, sorry. Oops. Sorry.
Speaker 5:Yeah. They're they're principles, George.
Speaker 1:Tom Tom.
Speaker 2:Go ahead.
Speaker 5:Go ahead, Arion.
Speaker 2:Tom. No. I'm thinking Tom. Stop.
Speaker 5:No. You you probably have
Speaker 6:a more interesting thing to talk about, so go for it.
Speaker 5:I know with VAST, their principal storage is QLC, and so they're the obtained is mostly as a right cache, so you don't wear out the QLC. So they they can pretty easily do that with enterprise SSD. Right.
Speaker 1:Interesting. Yeah. Alright. Good. So, like, very little changes.
Speaker 1:Maybe something, but very little. It's good. So they they the update is on. And I think that they were Tom, they were the biggest adopter of Optane that I was aware of. I just see who who was, like, the reference customer for Optane?
Speaker 5:Well, the killer app was supposedly SAP databases. Right. You could restart.
Speaker 1:Why am I forgetting about SAP? Of course. This is s a this is Adam's SAP's SAP HANA con conjecture.
Speaker 2:Yeah. I I don't know what specifically you're alluding to, but, like, if you put Octane and Hana together, like, my head is about to explode just in terms of things that I think are such pure bullshit. But I guess people have to play it hard.
Speaker 1:No. Your assertion was that SAP HANA HANA doesn't actually exist. And you asked to be proven wrong. Like, who has deployed it? SAP HANA is rolled out as the reason for every Intel feature.
Speaker 1:So they were using, I believe, secure enclaves. They were using transactional memory. Was it that was an SMP I mean, that now I actually don't know if I'm making it up or not, honestly. I I I and then because now as I'm hearing myself, I'm like, this is not that impossible. But I swear
Speaker 2:It's like like, isn't it obvious that, like, Hana doesn't exist? Like, you need you need you need to, like, stand, you know, on the mountain at midnight, you know, with the full moon for it
Speaker 6:to come into existence. For it to
Speaker 1:What look at it.
Speaker 5:Well, the other the other point of view was that SAP HANA was anything and everything that SAP was trying to do.
Speaker 1:Right. Which is so that's always dangerous because you've got, like, 2 hyped technologies are looking to one another to support themselves, like, that's definitely dangerous. I also feel like, look. Okay. Yeah.
Speaker 1:I thought SAP HANA exists. I don't have older siblings. Okay? The rest of you heard from your older siblings that SAP HANA doesn't exist. I wasn't I how was I supposed to figure that out?
Speaker 1:SAP HANA is your parents. Come on. Everybody knows that.
Speaker 5:Yeah. And it's it's kinda like 2 startups depending on each other. It's like it's not a big market.
Speaker 2:I've I've got a I've just handed this by Patrick Mooney also of Oxide that at General
Speaker 1:No. I like I'm sure Hana exists.
Speaker 2:Yeah. I'm I'm I'm sure it exists. I I choose to not believe it. Another one I choose not to believe.
Speaker 1:But so that was a I do feel that the and and this is where you do get into some very strange rhetoric from from Intel that doesn't really make sense about how Optane don't worry. Optane is being replaced by CXL. And have you heard this, Tom? Have you heard this particular line?
Speaker 5:Yeah. I mean, it it's kinda true that that I think at some point, the CPU guys kicked Optane out and said, look. Just put it on the other side of CXL. We don't wanna deal deal with it. And then it became it became even clearer.
Speaker 5:It couldn't really compete with the next generation SSD.
Speaker 1:I just love the idea of the CPU guys taking the opt in out. Like, it's a 20 something that refuses to get a job and can no longer live in the basement. It's
Speaker 6:Like, you're not supporting your wife.
Speaker 1:I'm like I'm just saying, like, I just I I love you, Octane, but you just need to get a job at this point. Like, it doesn't this is not
Speaker 2:working. Man, transistors aren't
Speaker 4:for Transistors.
Speaker 5:Because, you know, the the the early pricing of Apache Pass, it was only supported on the 4 and 8 way MP Processors, which themselves cost 4 to 8 times as much, you know, for for no chip benefit benefit.
Speaker 1:The okay. The Apache Pass was only on the 4 to 8 socket? Yeah. Holy. Yeah.
Speaker 5:In this initially, I don't I don't know if they changed that.
Speaker 1:Yeah.
Speaker 5:But that was how it came out, and so you had to pay a huge amount of money for CPUs to support. Okay.
Speaker 1:So the
Speaker 6:That must have been different because Facebook would never buy a Yeah. They they drew the line at this at this 2 socket system. They would definitely not have
Speaker 1:the mobile consumer to crack into the data center.
Speaker 2:Yep. Or at least a huge volume consumer.
Speaker 1:Right.
Speaker 2:Right. Like, it didn't it didn't need to be mobile. I mean, it's very very helpful that it was.
Speaker 5:I mean, these physics these physics things need
Speaker 1:Right. I
Speaker 5:yeah. To to get refinement. And and Intel was going after these super high margin applications where there's no volume.
Speaker 1:Right. And I think that it is a real struggle for the alternate physics to go after the high end. And I I Yeah. The it's it's a it's a struggle because it's gonna be very it has to be a very, very, very compelling use case economically. It is much easier 1 Yeah.
Speaker 1:Sorry. Go ahead.
Speaker 5:And any of these new technologies, you have to start with one very specific, very humble killer app and and then hope you can grow from there. But in Intel, I was, like, always like, well, it's memory and storage, and it's a whole new programming model, and you'll love it. And it's like, well, what's it what's it really good for today? There's never any
Speaker 1:Right. And what I also I mean, it's it's the the GPU GPU is the same way. Right? I mean, it's like it really needed the mainstream market to have the the ML applicability. I mean, I think it's like if if you do not have a GPU, that thing one did not need a GPU, would ML alone, maybe crypto alone based on NVIDIA's results today, would that have been enough to, like, to to drag it into being?
Speaker 1:I'm like, I'm not sure. I think it actually needed that that mainstream market.
Speaker 2:I mean, look at the SGI. Right? Like, our SGI
Speaker 6:got slaughtered mainstream markets.
Speaker 5:Yeah.
Speaker 2:SGI got slaughtered by from below by the ubiquity of the GPU, where then this high end thing just got, you know, eaten by the consumer thing that was able to reach that kind of volume.
Speaker 1:Right. So I think actually so it would be that the argument there would be like, hey, Optane, you've gotta find a consumer use case where it's not gonna work.
Speaker 2:Like, what's the consumer use case? Absolutely. But but you can also see how the message probably played out at Intel during its lost decade, right? Where nothing was really working, nothing was top of class, and here was some you know, unobtainium that that they alone had, and it was going to be the cure for everything. And when it reached flash level volumes, you know, the economics were gonna work with probably no questions asked about how it was going to get to those kinds of volumes or or what the apps were that were gonna drive it there.
Speaker 1:Yeah. That's interesting.
Speaker 5:You know, they they did they did have the consumer SSD product for, you know, laptop flash or cache management. I'm actually using one of those as a ZFS intent log now.
Speaker 2:You wait. You're using Optane right now?
Speaker 5:Yeah. You can buy 16 and 32 gig Optane SSDs on eBay for reasonable I SSDs SSDs in general are so freaking fast. Nobody needed caching for them.
Speaker 1:Yes. I mean, there's just not a compelling consumer use case for it, I don't think, over Flash. I Yeah.
Speaker 6:So would we would we use it as an always on type thing then? Like, in a let's say in a phone, like, you would never have to, quote, unquote, boot your OS. You would start it once. It would be in memory, and you would just keep it sitting there forever?
Speaker 5:Well, in theory, but, you know, they would have to do a lot of software work for that. So that would take a long time to bootstrap that.
Speaker 6:No. I agree. So you wanna do software updates and it's something you need to boot your phone anyway. So I don't know if there's actual value here. But that's, like, if you wanted to break into the consumer market, what would be the the feature that would make it the thing why everyone wants to dish dish flash and then go somewhere Yeah.
Speaker 6:Somewhere else.
Speaker 5:Well, no
Speaker 1:one Right.
Speaker 5:Nobody knows because because, you know, that that that that's the heart of the innovation is how do you really apply it.
Speaker 1:And it would be interesting, Ari, to know, like, okay. Let's just say, like, we've got something that is basically DRAM performance and it's nonvolatile. Let's just assume that. And and assume the economics deliver, is what that thing is able to deliver, is that compelling? And what would be is that compelling to the consumer?
Speaker 1:And I guess it would be like a phone that a phone and let's assume that our software does not data corruption, all the problems we talked about. Yeah. I guess it would be like a phone that you never Yeah. It's the non phone. I don't know if
Speaker 6:the power consumption, for example, would work out. Like, is it because it's nonvolatile? Does it consume virtually zero power when you when it's not in use? Can you turn it off more
Speaker 2:or less? Yeah. Can you, like, persist your memory leaks now? Yeah.
Speaker 6:Yes. Sure.
Speaker 5:Yeah. You definitely get some power saving, though. That would be nice.
Speaker 4:Do you get power savings on
Speaker 7:an awful lot of people
Speaker 1:obtained over I I mean, maybe or I I can't even get the I can't even get the question answered of, like, what is the draw of a DIMM, which is it turns out is like an unanswerable question. I don't know if you ever tried to
Speaker 5:this out. It's like, alright. Hey. Listen. I'm just like, you know No.
Speaker 5:I I I mean, I was assuming because it's not refreshing, but who knows what the basic draw is.
Speaker 6:But we still don't know how it works, so we wouldn't be able to tell whether or not it's it's actually power. Yeah.
Speaker 2:Well, no.
Speaker 9:This is
Speaker 3:a very good
Speaker 1:point, Ryan, because I I don't think we know how I mean, we again, it is phase change memory. I mean, it is or reRAM. It is gonna basically operate by putting more current through at some point to change the receptivity of it. But how much more? What does that mean?
Speaker 1:What does that mean for the actual, like and how does the draw vary on reads versus writes? I mean, I think there was a bunch of questions that we wanna get to that we weren't close to getting to because we were it was unavailable and uneconomic. We didn't know how it worked. But I think you would need to answer all those questions to know if there's something compelling there from a consumer perspective.
Speaker 2:Yeah. Matt, you got a question?
Speaker 10:Yes. So, well, a comment. So, you were talking earlier about the, the danger of persistent memory magnifying the effects of data corruption. And I find that, there I I only on my phone, I only reboot in 2 cases. 1, of course, is software updates, and the other is if there if if if there's something that's that's
Speaker 1:a miss
Speaker 10:a miss Right. In in the OS or and and and I need to just reset it. So so having having more persistent memory, as you said earlier, might be a misfeature.
Speaker 1:It might it
Speaker 2:it it works for it works for my best use case, which which happens all the time, which is me running out of out of battery. That's the most frequent reason for me rebooting my phone. So it would help out a lot.
Speaker 1:Yeah. My phone has this a the charming attribute that once it gets to 1%, if I plug it in, it consumes just enough additional draw that it will die as I plug it in, which is very nice. It's like, thank you very little phone. It's like and just for plugging me in, I will die. You're like, hey.
Speaker 1:Okay. Fine. Fine.
Speaker 2:Right. It's my fault for letting you get that low. And you're right. You're you're right to
Speaker 1:Exactly. You're right to find entry. But no, it's good point, Nava. We people we do we do have the idea that I mean, we we have volatility kind of baked into the way we think about these things in terms of, like, I can get this back to a state. I can't do persistent damage to this thing.
Speaker 1:And there is kind of a social contract with your software engineers that persistence is gonna be something done carefully, and we are not gonna generate something where it's bricked. But it's not clear to me that that would be, that would do a BU feature. And it's also not clear to me that it's I don't know if if if non volatility at DRAM performance speeds and economics. I mean, is it gonna happen? I don't know.
Speaker 1:I used to think it was for sure gonna happen, and now I'm a lot less convinced.
Speaker 6:Well, do you want it to happen is actually the question. Because we we kinda conclude that it's actually really hard to program.
Speaker 1:I think I do in that I was it the or this is where the CX helping does become interesting. So even though I think it's ridiculous that Intel is and at Tom, I agree with what you're saying about, like, it's Intel is kind of, like, half right that CXL
Speaker 2:Sorry. Sorry. Sorry. What is CXL? I was I was just Yeah.
Speaker 1:So CXL is the compute express link. This is basically, next gen interconnect,
Speaker 6:Nope. You're not making that up. That is
Speaker 1:It's like
Speaker 6:It's still
Speaker 1:But, like, how old do you have to be
Speaker 5:to errantly
Speaker 1:name a generation of interconnect after a literal generation?
Speaker 2:It's like, everyone stop. Pause. Pause. Pause. Did none of you have kids, like No.
Speaker 1:Exactly. None of you
Speaker 2:tell your kids what you work with and no. No. Alright.
Speaker 1:Exactly. And, also, it's like, I mean, Gen z, the jokes do kinda write themselves. This is like the TikTok interconnect or whatever. But the, of course, they did realize that, like, maybe this is a bad idea. So, Gen z ultimately, Gen z and CXO were rival standards.
Speaker 6:But they rolled it into 1
Speaker 1:They didn't roll it into 1
Speaker 3:of them.
Speaker 1:Yes. And so if you if you go to the Gen z website, you get basically a suicide note and how it's now, like
Speaker 5:But, but but Gen z was was 100% focused on load store memory semantics, whereas CXL is basically PCI plus plus.
Speaker 1:Right. And so yeah. Tom, do you wanna elaborate a little bit on what that means? Because I think it is I mean, I don't know. Maybe it's it's not as interesting.
Speaker 1:I I think it is, but I honestly
Speaker 5:Well, I think Gen z came out of the the machine from HP, you know, where everything was gonna be memory semantics. Are you serious? Are you somehow sending it.
Speaker 1:Just saying that for my benefit, or do you is that actually where it came from? Did Gen z actually come from HP's machine?
Speaker 5:I'm pretty sure it was all part of the same evil thing.
Speaker 1:Oh my god.
Speaker 2:We've come all the way back to
Speaker 1:Please tell me there's a Gen z ad somewhere in the walls of HPE that hasn't been not been released featuring Star Trek characters. That would be, just too delicious to hope for.
Speaker 5:But there there were a lot of people at both Dell and HP that were totally sold on Gen z. And they had they had they had some super nice physical layer stuff. Yeah. Very low latency. But, they should have applied that to something semantically useful.
Speaker 1:Yes. And it sounds like they I mean, that effort has been folded into CXL. So so, Adam, CXL will kind of replace my understanding. Tom, maybe you can correct me if I'm wrong, but it will effectively PCI ends up being effectively an implementation of CXL. So CXL ends up being backwards compatible enough with PCI that it's effectively, you know, one mode of transit is is over PCI.
Speaker 1:But it also will then allow us to treat things, as DRAM and in or or the way we treat DRAM today, namely coherent, and able to participate in in cache invalidation that is actually looking more like an IO device. So we can have CXL lanes to something that is an accelerator and actually have something that is like a GPGPU or another accelerator that's actually able to to participate, in the
Speaker 5:Well Yeah.
Speaker 1:Sorry. Go ahead.
Speaker 5:Brian, you're dangerously close to drinking the the CXL.
Speaker 3:I've had
Speaker 1:I think you're this coolie that I'm drinking right now. It's delicious. Where do where I was wondering where this came from. Good.
Speaker 5:Good. Because every everyone thinks that the c in CXL stands for coherent, but it it doesn't really allow any new caching capabilities. What it does is standardize the way that you do mass invalidates. And, basically, there's a, okay, it's your turn to cache things versus my turn to cache things on between CPU and device. Okay.
Speaker 5:So just But you could you could still already cache things over
Speaker 1:Just to show you that I really am checking the Kuwait over here, isn't that CXL2.overcxl3.0? Isn't c x it's like, hey, man. CXL 3 dot o solved that problem. I thought CXL 3 dot o is kind of has a Duke Nukem forever aspect to it in terms well, not that. That's overly pejorative.
Speaker 1:But it Yeah. It does feel that, like, isn't that isn't that what CXL 3 dot o was supposed to solve? My what I god. Get a load of me. What's going on?
Speaker 1:What what what happened to me?
Speaker 2:Well well
Speaker 5:well, we have
Speaker 2:Related question. How how real is any of this? Like, am I using this today and I don't know it? No. Are my children's children going to use
Speaker 6:this company? The corner.
Speaker 5:Just just a long one
Speaker 2:of your top ways. Right?
Speaker 5:C CXL is coming pretty fast. Yeah.
Speaker 3:It is.
Speaker 5:And, it it's got good things. You know, it's faster. But And
Speaker 2:and this is, like, some consortium of of, you know, everybody and their uncle in terms of Yes. You know, who who's doing this?
Speaker 10:Intel web.
Speaker 5:Well, it's basically in Intel Intel said, here's what we're doing, and we'll open it up for you. So everyone jumped on board.
Speaker 1:And a and, you know, with kind of the Gen z getting folded into it. So AMD has joined it as well. So we, like, we will we will absolutely be using CXL, Adam, at Oxide. CXL will be relevant for us. Now what what is unclear is whether that's c xlone.x or 2.0 or 3.0, which has different implications for performance, but also for the degree that these things can actually participate, in cache validation and so on and sending protocols.
Speaker 1:And then that to me has got ramifications for, I think, for and and, Tom, maybe god. I I just can't believe I'm, like, I'm I'm saying these things. But it does feel like it allows us to think differently about the way we compose compute elements in the system in terms of having these special purpose compute elements that are able to really participate, in the system in a way that right now, like, in NVIDIA g b GPU is part of this this other that you've gotta send things to. Am I drinking the kool aid over here? What's going on?
Speaker 5:Well, I'm not sure where they are in terms of the 3.0 spec, but I'm pretty sure with 2.0, it's still very primitive Okay. In terms of
Speaker 2:Brian, do you remember this at Sun? There was a you know, we had these multi socket systems and, I don't know if we designed it or Sun Labs designed it or some third party designed it, but there was like a, a video card, that you can plug into a CPU socket. Just Now it sounds like I'm dreaming.
Speaker 1:Yeah, exactly. That was the kind of the dream. And then what happened, Adam? And you were there. Yeah.
Speaker 2:But maybe you weren't there. This it sounds like this does not ring a bell.
Speaker 1:Well, so what rings a bell actually is something that's even nuttier where they put an MMU on a graphics card, and they forked the memory system in in the operating system. This is, for, Leo back in Sun Forum days. I don't know, Tom, if you had any overlap with Leo. This would have happened. I think I I think this is the kind of thing that Tom would have quit over, honestly.
Speaker 1:This is like the
Speaker 5:No. No. That this type of thing I wanted in from the early nineties was was for IO devices to be MMU compatible with the CPU.
Speaker 1:You didn't want it like this.
Speaker 5:Yeah. I don't know how I did it.
Speaker 1:It was it was a mess. It was a big mess. But that is not what you're talking about. You're talking about a graphics card that we yeah. That went in the CPU's.
Speaker 1:Oh, yeah. I feel like that. It that was not that's where I'm I'm I'm You're
Speaker 2:like, sure. That could have happened. Yeah. That might have
Speaker 4:been Yeah. That's right.
Speaker 2:Anyway, it's like Well, it Rather than, like, QPI or something, this is like a not fast, not as fast as not as kind of in the brainstem as QPI and, but a lot faster than PCI. Is that a fair summary of CXL?
Speaker 1:I would say incorporates the next gens of of PCI. And this is what allows for, I I'm gonna say disaggregation, but just your reminder of Jeff Rothschild. One of Jeff's rants that I really love is he hates the term disaggregation. He's like, you actually mean aggregation. You literally mean the opposite of what you're saying.
Speaker 1:It's not disaggregation. It's aggregation. And then Robert is is trying to coin amalgamation for this. But where you are taking like, this is where you kinda get to the bonkers future that I think Intel is talking about, where, oh, with CXL, you can now have a pool of future octane memory, whatever that is, that all of these nodes are able to connect to coherently. And that's, like, that's the kuwait I'm not drinking.
Speaker 1:I'm not I ain't drinking that kuwait. That that that that's where I draw the line. There are lines that I
Speaker 2:draw around
Speaker 5:here with
Speaker 1:Kuwait perspective. Yeah. I don't know about that one.
Speaker 2:I I'd like to see that one first.
Speaker 5:So if you want a little more divergence into Sun history, the the early nineties was the Wild West of networking where we're doing Ethernet, FTDI, and ATM, and fiber channel, and this thing called Taylor Taylor. Which was very popular with the hardware guys, and it was gonna be Coherent Shared Memory at Land Scale.
Speaker 1:Taylor like like Tinker, Taylor, Soldier, Spy? Taylor likes WIP?
Speaker 5:Yeah. T a y. Named after the efficiency expert guy, Taylor. But it it I could never explain to the hardware guys why this is such a bad idea.
Speaker 1:Someone help me Google this one. Are you able to are you getting
Speaker 5:No. No. It it it it was totally a sun clicker.
Speaker 1:Okay. Oh, this is it it it's gonna be internally.
Speaker 5:Yeah. But it it had a lot of support.
Speaker 4:Yeah. It was. A lot
Speaker 1:of people drinking that Kuwait? Yeah. Yeah. And I think I can get, like, the appeal of and I do think this gets into, like, why did Intel get so excited about Optane? I think I Tom, it was you or Adam that made the point that there was that it this was happening during Intel's lost decade, and Intel needed, like, a great hope on the horizon and obtained kind of so there's this kind of internal narrative inside of Intel that, like, Optane will save us all, which certainly didn't do anything to, like, to, to temperate expectations of the technology.
Speaker 5:Yeah. And I this the hype machine was so vivid. I I just was never I mean, I I don't think I was ever positive on the technology. Alright. Because they never showed me that one simple use case.
Speaker 1:Yes. I feel that I never got positive on the technology because they've never explained how it worked. And when I asked how it worked, they would say, like, you're the only person that cares about this. Nobody nobody else cares how it works. Only you.
Speaker 1:Like, oh, you're the only one that asks this question.
Speaker 2:Everyone that we've sold
Speaker 4:this to Exactly. Right. So Wait. Alright.
Speaker 2:How many people have you sold it to?
Speaker 1:And so I look. I kinda resent that whole line. I the I I really I people should not let their especially, you know, technology partners that they're gonna really, you know, go do some really deep things with. They should not be you should not tell someone that it doesn't matter that I'm the only one asking. It's not even true, personally.
Speaker 5:Intel had a had a corporate problem of being really good at training training their marketing people to lie. So the marketing people wouldn't know it, but they're still lying. So only place you could you could ever get the truth was from the register.
Speaker 4:Are you
Speaker 1:saying they had, like, some of, like, youth education camp where they would abduct future marketeers as, like, children so they don't actually know their Tom. They believe that they're telling the truth. They're gonna, like Pretty much. Around. Oh, okay.
Speaker 1:Alright. It got dark. Yeah. So it and in terms of the the those sort of kind of the lonely death of update I mean, I think the writing has been on the wall for this thing for a while, Adam. The this thing was not
Speaker 3:Yeah.
Speaker 2:Like, almost from its inception, it feels like, or at least from its first products. Like, each product was progressively less compelling as the expectations just got higher and higher. But the only time it seemed interesting was before it existed.
Speaker 1:Yeah. It kind of ended up like the it ended up like a gambler on a losing streak. It was like, we're gonna put more and more on the next generation. It's like, no. Like, I'm gonna make it all back in the next generation.
Speaker 1:It's like and it was just not happening. And it yeah. It is
Speaker 4:And it
Speaker 2:was like, oh, no. We're gonna we're gonna get the volumes by shipping the SSDs, and then that will bring down the price for the memory. And it's like, well, yeah. Right? You're never, like, you're just in hawk, like, you're never gonna make this back.
Speaker 2:You should just just pick up your chips and and go home.
Speaker 1:With what what few of your chips remain? And and then what's the business with Micron? The the the Micron marriage and then subsequent divorce over this thing is super weird. And, Tom, I don't know if you had Yeah. Do you have any insight into that?
Speaker 5:Not really, but but the real nail in the coffin was when Micron sold off the fab. That that was the only fab that
Speaker 2:Well, so like like I like I alluded to, I have I have a friend who's been close to this and has drunk way too much of the Kool Aid. And one of the things that clearly Intel has indoctrinated him in is, blaming Micron. Like, the the failure of like, no. The the reason why
Speaker 1:It's your mother. It's your mother. I I I actually wanna have this. I would be very reasonable. I just wanna see you every other weekend, and your mother's preventing
Speaker 2:it. It's like, Yvette, Optane would be fine had it not been for Micron losing faith, and then they'll never get to the volumes and and so on and so forth. So, I mean, he said this to me with a straight face like he believed it. So clearly, he had been hearing it from a lot of folks
Speaker 1:even though it's,
Speaker 2:like, total nonsense.
Speaker 1:And you and and you think that that that mister Shrub, our our our our shared acquaintance here, was a real advocate for Abkin. He had really he said he drank the
Speaker 3:Kool Aid.
Speaker 2:Well, drank the Kool Aid. Maybe it was just that he had so he had been working, with or or or like with folks from Intel on some collection of products, you know, in and around Optane. And and I sound vague, but that's that's as clear as he was with me as well. And, so I don't know that he he drank the Kool Aid necessarily, but I think that he was around he was around it so much and so many other folks were drinking the Kool Aid that some of these messages started to sink in and and take hold. And the other thing that thing that was funny is that, like, he found out about the death of Optane the way that, you know, everyone else did, That that no amount of proximity to himself included him in, which which you also get a sense of from the VAST blog post Yeah.
Speaker 2:Where, you know, they probably had not heard about this 6 months ago to to get their ducks in.
Speaker 1:I mean and they I mean, obviously, everyone knew that Optane had been transitioned to hospice care. I mean, this is not a huge surprise, I don't think. Wasn't it? I mean, I don't think that they, you know, they had done enough that that it was it was clearly struggling at at pretty basic levels. So it was not I don't know.
Speaker 1:I was not hugely surprised by it, especially after the the the Micron divorce.
Speaker 3:I don't
Speaker 5:know. I was I was working with otherwise smart people who were still pretty optimistic about Really?
Speaker 1:About Optane?
Speaker 5:Yeah. Within the last year.
Speaker 1:That's impressive. Do they would they could they volunteer how it worked? I mean, that's the I I I'm, like, looking for someone who just tell me definitively how it works. I wanna know I mean, again, obviously, it's PCM, but I or I should say, obviously. It it is apparently PCM, but I would like to know how it works.
Speaker 1:I see another couple of other folks here that, had joined us. I wanted to if anyone else wanted to jump in with their perspective on the the death of Optane, the the sad and lonely death of of Optane.
Speaker 7:Yeah. Yeah. So maybe I'm kind of biased on my opinion here because I've been working on Optane for almost a year.
Speaker 4:Oh, okay.
Speaker 7:Because that's, that's what I do research on. So I find, essentially, Optane to be a cool technology at the end of the day because, it kind of gives this bite of disability feature, although it's essentially a storage device and provides these huge increase in read performance, especially random read performance. So it was really hard for me to understand why you know, maybe, economically, it makes sense, but, I I I strongly believe there could be killer apps. I mean, if Intel allowed maybe a few more years to roll, that's my opinion because I clearly have seen or built applications that have seen 2 to 3 x performance improvements on Optane itself or viewing Optane as
Speaker 11:a storage medium. I I
Speaker 7:I totally understand that, the cost might not make any sense at all, but that's that's what I think.
Speaker 2:What I think that what I think those use cases might be, Vinay. Like, what what what are those killer what might potentially those killer apps have been that, and how and how would they get there in terms of the economics?
Speaker 7:Right. In terms of economics, I have no clue. Right? There's a I think CXL essentially, in my opinion, CXL killed Octane. Because now when PCI Egenfi is gonna give you such bandwidth capabilities that that what you get on Octane Demps, it doesn't really make sense to take that that precious area that your real estate on the dims.
Speaker 7:So I see it does sense to put it, behind the enterprise and note that CXL offers, store interfaces, I get on the dump slot. But as a as a as a technology in itself providing wider disability I mean, storage platform providing vital disability and increased random read performance. A lot of algorithms that, have been traditionally designed for, you know, doing traditional read, read accesses, traditional write accesses, which are all sequential, and also, you know, doing block kind of accesses increased a lot of read write amplifications. So all of that is gonna go away if it's bit addressable. So I find that super exciting, honestly.
Speaker 1:Okay. And the So
Speaker 5:I I have a contrarian opinion there, man. And, actually, the the thing I resent most about Intel and Optane is is how hard they pushed all the research community to do stuff. And presumably funded funded a lot of
Speaker 3:it,
Speaker 5:but basically a huge amount of wasted time. Whereas, what they should be doing is making memory virtual again. Because you could do all these things with virtual memory. It's just kinda slow these days. Why is it slow?
Speaker 5:Well, it's the CPU.
Speaker 3:Mhmm.
Speaker 8:So Okay. But isn't the
Speaker 7:What Isn't it the media also? I mean, sure, virtual CPU can be slow, but No. No. Media also
Speaker 5:It's it's it's page faults or just too damn slow. Mhmm.
Speaker 1:Okay. Page faults.
Speaker 3:We're
Speaker 5:we're getting as fast as as other things. We we'd be doing amazing things with virtual memory.
Speaker 1:And when you say page faults, you're talking about paging something in from a backing store, Tom. So this is like this is like a what we call, like, a hard fault where you are that you've got something that is missing in DRAM because it's been paged out to a backing store.
Speaker 4:Right.
Speaker 5:Which which is all that which is all that the obtained them was is is
Speaker 7:Well, not really.
Speaker 5:Right? Doing it in hardware.
Speaker 7:Not really because that's still slower. Right? Because if even if you did a page fault and you need to access something, if it's not obtained, you still need to do a 44 KB block access. But if it's a byte addressable device, you don't need to do that original amount
Speaker 5:of Optane obtained used the Optane DIMMs use 128 byte cache lines.
Speaker 7:Right. Right. Essentially, it was
Speaker 5:So that that was the minimum. So so what about an architecture that had variable size variable size pages where you could have small pages when you need them?
Speaker 8:Mhmm.
Speaker 7:Yeah. Sure. So that that's my point again. Right? So you it's it's a good idea to provide byte accessibility feature through the hardware itself rather than having a virtualization layer that will sub that will kind of give you a a a shutter of biter disability because at the end of the behind the scenes, you will still be doing a large amount of, data transfers.
Speaker 7:Like, you would want to avoid that data moment unnecessarily. The whole point of opting or, like, bit of usable storage is to be efficient in accesses. Right?
Speaker 5:Okay. So Yeah. I'm I'm I'm gonna make up a bunch of red hats and say, make memory virtual again.
Speaker 8:Well and and so one one of the other things that I find very interesting across the board is that everyone thinks that obtained is a great solution to a problem they themselves do not have. And and it Yeah. Like, academia looks at it and says, this is a really great solution because it's got these xyz factors, and it might be good to do this. And the people in industry who are doing this look at it and say, no. It costs too much.
Speaker 8:It or it it doesn't actually behave well according to x y z real world conditions. And, like, you know, the the media obviously looks at it and says, yeah. This looks like it ought to be good for something. But there there's a pretty decent number of people on this call who have all looked at Optane and said, yeah. You know, it's probably good for something, but not me.
Speaker 8:And with the exception of VAST and HANA, I I don't think I can really name many people who said, yeah. It's a good solution for me.
Speaker 1:Well and then Matt, I think it's a good point. And then you also think, like, alright. If that's the problem you've got, is this the best solution? Is this the most economic solution? I mean, I I think, I Vinay, I think, first of all, thank you very much for joining and it's great to get your perspective.
Speaker 1:I I won't let Tom beat you up too much for being, you know, I that Tom's treating you like an Intel marketing arm. You know, we've got a I think it's great that it's a I I don't care if your research is funded by Intel. I actually think it's good that Intel is funding research.
Speaker 7:Also, honestly, I don't think I don't think I'm funded by by the way. It's being honest here.
Speaker 1:Alright. Well, that there was a it's it's okay if you are. It's fine. It it it doesn't, but the I I, I do think that the economics and I mean, this is something that I do think is important, is that economics is very important in in our industry, and it's a very important dimension of computer architecture, and it's tough to factor that out. Certainly, we can't factor that out.
Speaker 1:I would also say that power is something increasingly that we can't factor out. I think power has been often factored out, and you gotta you know, as we as we know, at Oxide, I mean, I think, like, I I I like the idea of smartNICs, for example. But smartNICs come at a a a power consumption a rate of power consumption that is really problematic when you start deploying racks and racks in them. So I I actually think you've got you've gotta you can't factor that stuff completely out. I also wonder so you said it's, like, 2 to 3 x for random read workloads.
Speaker 1:And, presumably, you're looking at the you've got workloads that exceed the working set of DRAM but are small enough to fit in Optane. Is that a fair summary?
Speaker 7:Right. Right. So, you now I think maybe this might not be a realistic setting where, in one in one configuration, I viewed Optane as the final storage medium that you would wanna reside the data on. Right? Like, if you you have that surface on top of obtained, the that's the, essentially, the file system in the app direct mode that you would configure obtained in, and you would treat all the data that that there is, that's there on on that as the final destination.
Speaker 7:So it's not as if whether of is the data going, is if the data is not gonna fit in in the DRAM, then it's gonna fit into Optane, but it's, it's not about data that's not gonna fit into Optane. You're essentially right. Like, what whatever you said is is what I looked into.
Speaker 8:Right. And I've seen, I think, a terabyte obtained DIMM versus the 256 gig DDR 4 r DIMM. Right? Mhmm. They they make terabytes now.
Speaker 8:Right? Mhmm. So so you're talking at most of 4 x increase in capacity at basically the price of DRAM with That's right. Price per gig of DRAM. Mhmm.
Speaker 1:Is it the price isn't it or more expensive than DRAM, right, from a price predict a perspective?
Speaker 8:You know, depending on how you wanna fudge the numbers around for them. Yeah.
Speaker 1:I mean, certainly by the end, I think Intel was and actually, Intel is sitting on a bunch of inventory. That's this is the real question. It's what happens. Now that the now that this the the kind of, like, this thing is this body has kinda toppled over on the Silicon Valley Serengeti, who are the scavengers that come and pick up the opt in inventory? What happens to that?
Speaker 7:Well, I I think it comes down to what you were talking about, Brian. I mean, no nobody builds, I mean, very few people build large scale memory systems nowadays. Right? It's all, it's usually you span out and don't don't scale up. So this is a very nice solution for scaling up, especially when you want higher memory density.
Speaker 7:When you, when you wanna populate or buy larger, like, let's say, 512 GB DDR 4 memory, it's gonna be very costly to build this kind of machine. Right? But if you have a mix of DDR 4 and Optane, it should be much cheaper. It's what the presumption that Intel
Speaker 2:Dene, you're but we're still in denial here. You're you're you're
Speaker 4:talking about because
Speaker 2:you're talking about Optane like it's gonna be something that you
Speaker 8:can buy. Right. Right.
Speaker 1:Adam, let him mourn. Let him mourn in his own way.
Speaker 6:Yeah. Exactly.
Speaker 2:This is an intervention. Don't know. Morning.
Speaker 1:Well, I know
Speaker 2:we will let this all
Speaker 8:so know. Building it out of Optane is actually any cheaper than building it out of DRAM is, which is really, like, the big killer. Right? Right. Like, you know?
Speaker 1:I I kinda like the fact that Dene has brought his snowing salts to the funeral and is trying to revive Octane. I think there's a but the I I I feel like there is something charming in it.
Speaker 8:Like and and I I will I will agree it that Optane as a underlying technology would have been really interesting. I mean, I I came to it more from the Micron side than the Intel side. Like, I've I talked to Micron in the past about, like, hey. You know, can we just buy those chips? Because, like, I'm I'm not running Zions here.
Speaker 8:I'm I'm building, you know
Speaker 1:Totally. Okay. So, Matt, I'm dying to
Speaker 8:ask you. Performance. Yeah.
Speaker 1:Did did mom shit on dad the same way that dad shit on mom? I mean, how did they
Speaker 8:Well well, so in in Micron's words, Intel threatened to sue us if we sold it ourselves.
Speaker 1:Okay. So so this this this this marriage is going south.
Speaker 8:This marriage You know? Like, they they they basically said that they had, like, Intel put in most of the r and d money on it, and therefore claimed most of the IP rights on it. And therefore, the pricing from Micron would have been even worse than the pricing from Intel. Because, remember, Intel was selling this at a loss right up till the end.
Speaker 5:Right.
Speaker 8:I mean Like like, they they were like, yeah. I mean, we we could, in theory, try to make a commercial product and sell it, and we'd sell, like, 2 of them. You know?
Speaker 1:Right. And so, I mean, they were trying to make the economics I mean, they were trying to defy the economics. Clearly, like, the and if it would be helpful to know how it actually worked to understand I assume that, like, part of the problem is just in the actual fabrication of this. So it is I
Speaker 2:have some interesting I have some interesting notes on that.
Speaker 1:Yeah. Yeah. Go for
Speaker 2:So, from from our man on the inside
Speaker 1:Mister Shrirad.
Speaker 2:He he is, who's who would love to join, but can't, tells me that there's a warehouse full of, Optane, both raw SSD and in the DIMM form factor just sitting there. And the plan is to to draw out of that and manufacture things out of that, until the EOL and doesn't know the data that he refers to this as the strategic opt in reserve.
Speaker 1:Oh, the I don't I'm getting a little concerned because we've already described him as, like, drinking a little bit of the Kool Aid. And, I mean, clearly, that's a Jonestown reference. This is just this I I like, where you're going feels like the I is there a rapture that's gonna come? I I feel like there's See. Is judging day coming?
Speaker 1:Is update going to judge the living and the dead? I just I'm I'm concerned.
Speaker 2:So apparently, one of the big challenges of update
Speaker 8:wasn't will rise from the dead after 30 days.
Speaker 1:There you go.
Speaker 2:So, apparently, one of the big challenges with Optane was with the fab. And one of the differences between Optane and Flash and isn't the details, was that like Flash could be switched out for, you know, MLC or TLC or SLC or QLC, with without a huge amount of overhead. Whereas, Octane, it, you know, based for reasons he doesn't fully understand, they basically had to dedicate the fab to this. And so what they what they did recently is did a big run of Optane. That is the last Optane that will that that we, humanity, will ever have.
Speaker 2:And it's sitting around in various states of completion.
Speaker 8:So, Adam, I I would also add to that that it's not just that you can switch the fab between SLC and TLC and 2LC and QLC. You can actually switch chips at runtime on a page level between those different modes if they're higher level. And and you can trade reliability for capacity in real time, which is something that I've I mean, who knows with Intel? Like, Intel clearly, Optane can do it because it's magic magic memory. But, you know
Speaker 1:And so, Matt, are you saying that because I do think this is kind of an interesting point about, like, how I mean, a traditional way if you're in semiconductor manufacturing, you've got ways to change yield. You can bend things. You I mean, the end it's been done for time memorial where you are able to take parts that don't that aren't performing as well and then those get binned differently. Are you implying that Optane doesn't have perhaps some of those advantages and therefore, economically, it's just they've gotta they don't have this ability to, kind of dynamically change the yield, which makes sense. Again, I don't know.
Speaker 8:I I mean, I think I I can since no one knows how it works, I'm I'm somewhat hypothesizing here. Right? Right. I think I think what happened was basically, you know, like, with typical flash memory, you'll have a whole bunch of pages that get lumped together in a chip, and you get to, you know what like, flash memory doesn't come with a statement saying this is a 2 terabyte chip. It comes with the statement this is a 2.1 terabyte chip.
Speaker 8:And by the way, there's, you know, 5% of this isn't bad. Figure it out for yourself. It's left as an exercise to the reader. But by the way, we guarantee page 0 is good, so at least you know where to store your table of all the other people who are bad. Right?
Speaker 8:And you can go through and you can change your error correction. You can do other things to to try to keep this, you know, keep this whole circus running. I think Optane, what they ended up doing, just given the fact that there's a lot of mysticism and there's a this in highly intelligent controller and there is this DRAM cache on DIMM. I think what they basically ended up doing was stripping away a 100% of everything they had planned for process yield and basically saying, well, we think we can make this work if we fix it all in software.
Speaker 1:Oh, god.
Speaker 8:And and so, basically, like, I if if you are having yield issues, which it sounds like from all sources they were, like, my my best guess is they're they, like it sounds kind of insane.
Speaker 1:Nope. Not sure.
Speaker 2:Adam, do you Yes. You were you were there, but Matt seems to be
Speaker 1:gone. Okay.
Speaker 5:I I So I think where where he's going is is that the Octane is not going away. It's just being frozen for future generations to revive.
Speaker 1:It's been cryogenically frozen. So hey. There because some other folks at I'd, West and and Sidar would want to get you give you opportunity to to hop in here.
Speaker 5:Yeah. Thanks, Brian.
Speaker 12:I was at IBM when all this was going down, and, one thing that IBM was interested in for years was putting, PCM on DIMMs. And IBM had already moved the memory controller out to the DIMM. So each DIMM could have its own memory controller that was specialized for what was on it. So you could have, you know, a DRAM memory controller, you could have a PCM DM with a PCM memory controller. So when Optane came out, IBM was like, great.
Speaker 12:Let's it's obviously PCM. Let's buy it. And we heard from Micron that, like, oh yeah, the chips that we sell are only for storage. You can't use them for DIMMs. Because and somebody else already said this, basically.
Speaker 12:You know, Intel basically owns the idea of the DIMMs.
Speaker 5:Oh, boy.
Speaker 12:And, you know, they wanted that Xeon lock in, so they didn't want, like, even IBM to be able to make, you know lock in.
Speaker 5:Intel is always doing things with their new tech.
Speaker 1:Intel, Come on.
Speaker 12:Yeah.
Speaker 5:They put so many marketing restrictions on things that they don't let other people forgot how to use them.
Speaker 1:But, like, talk about not dancing with what brought you. I mean, that's not what made Intel the company that it is.
Speaker 12:So I I think Micron was almost like a fake partner. That's how I think about it. That it seems like Intel did all the work. You know, Intel actually developed it. Intel was the only customer for it.
Speaker 12:And Micron was, like, a a third wheel or something that was just not needed.
Speaker 1:Micron was Intel's beard for, just to go home with the family?
Speaker 12:So Yeah. Yeah.
Speaker 1:I mean, do they need a manufacturing part? Or what was the, I mean, what was the genesis?
Speaker 12:Yeah. I don't I have no idea.
Speaker 9:So I think
Speaker 3:the Please
Speaker 1:write a book about all this so we can read it.
Speaker 9:So I I I was gonna say, I I I think the fab, from the Micron side was probably what Intel really needed because they didn't wanna cut down, capacity in their main thing. And they already had, you know, been dinged hugely for their, logic stuff around that time. Right? So that's a. But, b, I actually don't think, generally, PCM or any of the, quote, unquote, memory stores are dead.
Speaker 9:Although, caveat, you know, here that this is also a lot of what my research is on, so I am biased. But, give I mean, I think IBM's been showing PCM down at 14 nanometers. TSMC is offering resistant RAM in commercial now, so not even risk. So I'm not quite sure that they're dead as a general concept. They might be dead, to obtain specifically.
Speaker 1:I mean, I hope not, honestly. I mean, I actually I I just think just to kinda end maybe this is gonna bring us around to where we started, but I'd I I think it's really important that we are actively engaged in researching alternatives and and that we some of these alternatives that are showing promise like PCM, you know, that's been around for a long time, but, like, continuing to develop that. I think it's really important because I think we we it's just important to kinda develop this in the right way. We can't lose track of the economics. But I I think I'm glad that that I guess, I don't I don't view PCM as being necessarily dead.
Speaker 1:I I view Optane's embodiment of PCM despite whatever whatever our, mister shrub thinks about the the the coming the coming rapture of Optane memory. So, Adam, what happens with the warehouse of Optane? The the the
Speaker 2:They they draw down to for, for, like, replacements. So, like, they've got some contracts in place. They've got some parts under warranty. And I guess this this sort of Indiana Jones like warehouse just just gets drawn down as it's needed, until they finally say, you know, we're we're out, or they they take what's left and throw it out back.
Speaker 9:So, actually, a quick question. Has anyone used Optane systems, like, amongst you all? So have you had user experience code on code from that side?
Speaker 1:So, I mean, the Arian did. I did not we did not because it was never it it we just couldn't make it economically pencil out. And then also the fact that they wouldn't describe how it worked, and then it was kind of, like, constantly shifting about what the focus was. So I think we and by we, I mean, at originally at Joint and now at Oxide, I think we're probably like a lot of folks. It's like, definitely interested, definitely curious, but not in this embodiment.
Speaker 1:And so, no, we did we did. Adam, you you never used it. Right?
Speaker 2:I mean, not not in anger. I, like, benchmarked the drive, really out of curiosity, and it was fast. But for all the reasons that you just described, it didn't go beyond that.
Speaker 9:Yeah. Because I kinda wonder if it went down the same way, you know, the PS 3 cell did. It it the performance was amazing, but then expensive to make. And then just the programming model was so bad that, you
Speaker 7:know, no one really wanted to move.
Speaker 9:So that might because, honestly, that's one of those things that, I think ends up killing a lot of these products. It's not so much I mean, Intellanal, regardless of what anyone else says, their hardware is good. Like, they have really clever people. Right? I think the problem ends up being once you go up the abstraction there, like, it's just really difficult to think about what kind of correct programming model you should use for one of these memories.
Speaker 9:Okay. That can end up being a major issue.
Speaker 5:I think that by pushing a new programming model, they hurt themselves. They they should they should
Speaker 1:have Yeah.
Speaker 7:Yeah. Yeah.
Speaker 5:Said, this is really useful before you do any software, but most people couldn't couldn't do that.
Speaker 1:I think if it had been really compelling and if it had been economically compelling, we would have dealt with it in software. We would have found a way. And it may have been a new abstraction, but I think the sell comparison is interesting. I disagree with it, but it's interesting because the I I mean, cell had a super wacky programming model, and I and I know that, Jonathan was on here earlier, and we went into that in our our episode with Jonathan Blow talking about Cell a lot, which is really interesting on on on metal. But the I don't think that this is so radically different that it couldn't have been programmed around.
Speaker 1:And I think if it economically delivered and certainly we've seen this. The things that are indisputable. Things like, you know, SMP. I mean, SMP changed things radically for software, but it also was economically indisputable, so we dealt with it. Things like caching.
Speaker 1:Right? Caching caching made software system software are, like, quite a bit more complicated. But economically, it was indisputable, so we dealt with it. So I feel that if it had been if it had delivered on the performance of the economics, we would have dealt with it. No question.
Speaker 1:I I think that we, I saw I'd so I think it's an interesting kind of proposition, but, I and certainly the fact that to the degree that it didn't require any, you know, a a change in the programming model, that didn't help. I mean, one question is, like, does the nonvolatility even matter? Certainly for Feneus' case, the nonvolatility wasn't all that important because it's just, like, dims effectively.
Speaker 5:But but, Brian, I I can argue that for SMP, it took decades to sort out the the kernel.
Speaker 1:It did. It it it took it took no. It took it didn't take decades. I mean, it was a mess a little bit, but it also was like it it it it was a mess, but it was also economically indisputable. I mean, there was, like, the the there was just no way you were gonna get away with being a unit processor operating system.
Speaker 1:Folks tried. Right? There were there were SMP deniers. Right? Like like, did HP UX famously have the entire optics to run on one core?
Speaker 1:They or one CPU? They ran only applications. It's like, obviously, that's just not gonna work. And you the ultimately, everybody had to adopt SMP, and there were people that believed you couldn't make system scale, and you can. Because the economics always win.
Speaker 1:It's kinda it's it's my feeling.
Speaker 13:It's still still kind of a mess, though. Like, what's the right argument to put after make dash j? Fair. Your OS still has no idea.
Speaker 1:Yeah. So I mean okay. I fair. Although, I think that that's for other reasons. I do think that, like, the the opacity of the, of the system in terms, that makes that that's harder than it should otherwise be.
Speaker 1:But we are you are able to run a a big parallel make, and you are able to leverage all those scores. And we and we're not going back to that. Right? And we're not we're not going back from PPGPUs either.
Speaker 13:Yeah. It's true. It's just if there's 2 of you, you will you will fight to the death over those cores if you both try to run that parallel make together and your OS will have no idea what to do about it.
Speaker 1:That is true. That is true. But it's still yeah.
Speaker 7:I mean, there are still a simple abstraction as well available on Optane to honestly speaking. Right? There is still a file system that you can put on those obtained DIMMs, and you can map a file and use it as a biodegradable, malloced area. So, like, I agree with you, Brian, that if it made economic
Speaker 1:did Vinay go away for Did did did did did did did did did did did did did did did did it just me? Oh. No? Alright.
Speaker 7:Oh, am I audible?
Speaker 1:Yeah. It's Twitter spaces. I I if if only Optane could have solved the Twitter spaces problem. I I Mark, I think we're gonna
Speaker 7:Can we hear you now?
Speaker 1:Yeah. We can hear you now. Yes.
Speaker 7:Okay. Perfect. So, yeah, like like I was saying, there was there was always an a simple abstraction on obtained available. Like, for example, you could just put a file system on these obtained terms and then, memory map a file and use it as a malloc area to do all your interactions with it. Right?
Speaker 7:So I I agree with you, Brian, that there is always a there's always ways if if it's if it economically makes sense in the software layer.
Speaker 1:Yes. If an economic makes sense, we're gonna deal with it. I mean, look, like, no one's on x86 by choice. Jesus. I mean, we we would be on a different instruction set architecture.
Speaker 1:Right? We we would robot a pain for to to accommodate x86. Alright, Mark. I think we're gonna let you get the last word in here to kinda, wrap up. What's your, welcome to the Optane Funeral.
Speaker 1:What are your,
Speaker 11:Yeah. I'm truly sad about it, actually. I actually am using Optane in production. It's about twice as fast as NAND flash when it comes to pulling data sequentially off of it, and about 4 times faster at random 4 k reads. But, overall, given the price of these obtained drives, actually just not worth it because they're basically 10 times the cost for maybe 2 times to 4 times the performance.
Speaker 1:Yeah. Yeah. That is what that's a just a gutting note to end on. I mean, because that's it. Right, Mark?
Speaker 1:I mean, that is basically that's kind of the beginning and end of it. It's like, hey. I've I'm you're using it in production. The technology works at some level. It's just that economically, it's like juices and worth the squeeze.
Speaker 2:Where whereas NAND was the opposite ratio. It was like, like, you know, 10 times faster and and, you know, 2 x the price. Not not precisely, but it was more of those kinds of orders of magnitude or more of those kinds of scale than than this inverse.
Speaker 11:I was really hoping the economy of the scale was gonna fix that, but, apparently, that's not ever gonna be the case. So I'm I'm not doing anything special in my rust code to say that this is an obtained drive that you're writing to. So I'm probably leaving some performance on the table. But even so, like, it's a it's a big hill to climb when it comes to 1,500,000 IOPS on a single drive is really nice.
Speaker 1:And then, Mark, I I assume you have an application where you the the kind of the the dataset that you're working with exceeds DRAM and fits nicely in that the the the span of the offline drive.
Speaker 11:I don't know. Actually, I needed it specifically because I needed the storage to be absolutely bomb proof. And so when I write to the drive, I wanna make sure that when it's written, when the OS says it's written, it actually is truly written out and not hitting a DRAM cache before the flash. So I wanted to make sure that it absolutely was on storage, when it when I get back the call from the OS saying that, yes, this is written.
Speaker 1:So you actually are using this for its write performance. And and that's
Speaker 5:the I'm
Speaker 11:using it for yeah. Absolutely. I mean, I'm using it for, an SQL light database of all things. Basically, Wendell Wilson of level 1 text, put me onto the drives after saying, like, you can see pretty significant performance in this one specific case when you're running to an SQL lite database where it will be at least twice as fast as what you are before.
Speaker 1:That's and you're able to have all the Internet hang off of a single SQLite SQLite database or whatever it is. I mean, v v you've got a use case that really needs that performance, and it was able to to have that performance shine.
Speaker 11:Yeah. And and that's, like, the only reason why I kept the drives at this point because I can't really guarantee it properly without doing a bunch more programming work to make sure it works somewhere else.
Speaker 1:And okay. Then I've gotta ask you though that just the limitation with respect to Intel CPUs, does that bother? Because you're obviously having to use, you know, isolation. Yeah. I'm still in
Speaker 11:the EPYC?
Speaker 4:Well I'm running on a Nissan chip.
Speaker 2:Because it's not like it it's new to it's Oh, because you're using
Speaker 1:the drive. You're using the drive. Yeah. Yeah.
Speaker 5:Right. Right.
Speaker 11:I'm using a p 58100 x.
Speaker 5:Right. You're using the right.
Speaker 1:And so has the DIMM form factor ever been tempting or no? Because of the you're using Milan. So, no, the didn't form factor is not the tempting.
Speaker 11:Been tempting for me. I mean, my dataset is not big. My entire application will live in 1 gig of memory. And I keep on, like, installing Debian, and it's, like, keeps on telling me, hey. You need swap.
Speaker 11:Like, I definitely don't need swap. You don't have to trust me on this. I just need right moments.
Speaker 1:Right. Right. Well, I feel that this is a fitting end. I I mean, Mark, I feel that, like, you that's a fitting final word. It's like, hey.
Speaker 1:Economics are really, really important in architecture, and this thing was interesting and promising. And the underlying technology in terms of PCM is not dead. And we may see alternative physics that that ultimately break the economic barrier, but it ain't Optane.
Speaker 11:Yeah. I wanted it to be good, but it just it's not worth it.
Speaker 1:It's not worth it. Well, rest in peace, Octane. Roll on Octane Rapture for those who are waiting for the warehouse of Octane to come again and judge the living and the dead. And, yeah, Adam, any final thoughts?
Speaker 3:No. I
Speaker 2:mean, always a, you know, a a a solution in search of a problem that it never quite found.
Speaker 1:And there we go. Alright. Well, hey. Thanks, everyone, for joining. Hey.
Speaker 1:In a in a very, an an uncharacteristic act of advanced planning, we actually know what our next two Twitter spaces are gonna be. So we're gonna do, next week, we are very excited to have Kate Hicks from the Oxide team talking supply chain. So, Adam, I'm really looking forward to this one.
Speaker 2:Yeah. It's been great.
Speaker 1:So, you know, it's a question that's come up for people asking what we're doing at Oxide about the supply chain. Kate, who was at at Apple and actually Intel briefly before coming to Oxide, has been instrumental in what we've done about the supply chain. So really looking forward to getting Kate's perspective. That's gonna be a lot of fun. And then the week after, we're gonna have, Eric on the engineering team that you've heard from Tales in the Bring Up Lab, along with a couple of other folks, talking about bringing up our 100 gig backplane, and all of the adventures there.
Speaker 1:We used to be, you know, oxide, needed a 499 ohm resistor. And to live, if you if you had listened closely to our tales from the bring up lab or more tales from the bring up lab, and that has been bested by the 1 100 ohm resistor that we needed to get the cable back plate to work properly. But, we're gonna talk all about that and about signal integrity. It should be a great Twitter space, and that's coming up in 2 weeks. So join us then.
Speaker 1:Thanks as always. Thank you everyone for for, for participating. Those of you who are actually running SAP HANA, we're pretty sure it's not actually your parents. So don't know. That's actually
Speaker 2:it's not
Speaker 1:a trade. Alright. Thanks, everybody. Take
Speaker 2:care, everyone.
Creators and Guests
