SPEAKER_01 0:00

Marvell, everybody, the next one trillion dollar company. And I was like, ha ha ha, funny joke. And then I learned the next day, or I don't know, the same day, whenever the markets opened, it's like Marvell is up 70%. And I was like, oh my god, there was my arbitrage event. And I was there taking Jensen like the jacket video the whole time. My video cost me thousands of dollars. I should have bought wow, I saw a hundred billion sentence come by right by my face, and I just let it go by. Like I did absolutely nothing.

SPEAKER_00 0:36

All right, hello everyone. Welcome to another semi-doke podcast. I'm Austin Lyons of Chipstrat, and with me is Vic Shaker from Vicks Newsletters. So we're back from Taiwan, and uh I'm getting over my jet leg. This is my first time to Asia, and as hard as I tried to not have jet leg, it's still just my internal clock is just totally screwed up. Um so I'm back, I'm awake, I've got some coffee in me. And listeners, um, you may not know this, but Vic and I, we've got we're six months into a podcast, um, but we're actually in two different countries, and we actually met for the in person for the very first time. So Vic, you took a picture, it was pretty cool.

SPEAKER_01 1:15

Yeah, it was awesome. It was totally cool to meet you. And when the you know, when you s when we meet each other, the first thing that came to my mind is like, holy crap, Austin is tall, man. What you don't like in video, I can't really tell, you know. Sure. And uh when I met you in person, I'm like, who is this guy? Like, why is he all the way up there? That's hilarious. I'm like this small Indian dude. Uh so I'm like, yeah, this is this is like my regular height. I'm I'm I'm kind of like average, I would say, or for Indian height. But uh yeah, Austin's tall. So that's cool. Uh so we took a picture, right? So yeah, obviously you can see the height difference there. We are on the same floor level. That was that is cool. Yeah, it's nice to meet you in person, finally. You know, after all this talking online, it's amazing what you can do with the internet these days, right? You can start a whole podcast. We got this thing, we talk so much semis, but in person it was awesome. Like we met so many people that were also in the online world, and you meet them all in real life. Computex is such a big show. It's amazing how big it was and how many people there were. All of them looking at this AI hardware stuff. I'm like, wait, are all of you really interested in AI hardware?

SPEAKER_00 2:26

I'm amazed. Yeah, totally, totally. Yeah, I agree with you. It's really cool to meet, obviously, you in person and hang out more. And then uh just the like other sub stackers, YouTubers, uh, people in the industry that we know of. Um, it's very humanizing. You know, you just get like in face to face, and you're just like, yeah, this is another person who has the same crazy interest that I do, going really deep on the businesses and the technology. And and yeah, I agree with you. On the one hand, like there was a ton of people there. And on the one hand, I was like, well, you know, we are in Asia and a lot of people work in this industry, but on the other hand, it does, outside of the whole like AI and semis has blown up in the last two years, it is still a little crazy. Like, yeah, I've always been into this stuff, and it's pretty like kind of deeper tech than maybe just pure software. And it's crazy to see, like, oh, there are hundreds of thousands of people that also feel the same way. Of course, as they're walking through, I thought, I wonder how many of them have seen the podcast. And if not, there's a huge subscriber base right here, like in the Computex uh pre-registration line, where apparently I actually didn't pre-register. So I there was like literally a thousand people in this line where if you pre-registered, you got to pick up your badge. And somehow I screwed it up and didn't actually pre-register. So I got to like go to a different place where there's only 10 other people. And I was like, okay, I guess I'm never gonna pre-register because I got like the fast pass. But I looked at that long line and I just thought, like, we got to get all of them to check out semi-toped.

SPEAKER_01 3:54

Like Austin is over there counting Tam, you know, like, yeah, there's my TAM right here. Computex is my TAM, my Substack TAM, my podcast Tam. But yeah, you're right. That's right. I'm like, I was truly amazed. And uh yeah, it's so funny though, because in the previous evening, I think you had some events you had to attend. And by the time you showed up and we met, it was like uh already like at four, four past four p.m. And uh like we went to the registration counter and they were like, no, it's done for the day. And the next morning I walk into the venue and I see the lines on the registration, it goes like the entire length of the conference hall or something, like you know, yes, half a mile. I'm like, oh my god, like Austin should have done this yesterday. He just needed to show up like 10 minutes earlier, and now now I don't know how he's gonna get his badge. But I'm glad you're out there counting Tam, you know.

SPEAKER_00 4:44

Yes, totally. Yeah, I thought the same thing. I was like, oh crap, I screwed this up, but then it turned out okay.

SPEAKER_01 4:49

You know, another thing that stood out to me was like there are so many people who came up to us and said, like, hey, I'll watch your podcast and like I like I like it, and I love what you guys do and all. Super thankful for meeting these people in person because it's it's insane. Like, we don't know who's on the other side of this video, right? But there are there are people apparently, and it's nice to meet them.

SPEAKER_00 5:08

Totally, yes, yes. I I felt the same way. Uh, you know, when people subscribe to our Substack and read it, you know, we can see the email of who signed up. So you get a general sense of like, you know, who your audience is. But with the podcast, we just kind of throw it out into the ether. And so it's cool, yes, when people came up and they're like, oh, hey, Austin, what's up? Um, yeah, I had someone at breakfast one day that was like, Oh, are you, Austin? I was like, Yeah, what's up? And then, and then he was like, Oh, I think I you do you write a substack? And I was like, Yes. And then I was like, I have a podcast too, you know, but just cool to run. It's like um, it's like niche maxing or something, like in a very, very, very narrow niche, people like know who we are, and that's very cool. But on my flight home, one my flight got, I was supposed to fly through Tokyo, but that got canceled because there was a typhoon. So I hope all of our Tokyo listeners are okay, Japan listeners. And and so I took a different flight, Taipei to uh SFO, and when I got there, someone sat down next to me. I was like, oh, that looks just like Ben Thompson. And so then later I was like, I think that is Ben Thompson. So I said to he's str he writes Stratectory for those of you who don't know. And so then later I was like, hey Ben, I'm Austin, you know, and I introduced myself to him. But same, I thought about him. I was like, oh, he's like hyper famous in a certain slice of the world, and then everyone else at the airport has no idea who he is.

SPEAKER_01 6:24

Yeah, he's uh for those who don't know, I think we should mention this. Uh, Ben Thompson is the OG newsletter guy. He started a newsletter when there was no Substack or anything like that. Substack is built on Ben Thompson's idea of creating an email newsletter. And, you know, he has tens of thousands of readers and he writes every day. And, you know, he has a podcast. So we we all know him, you know, in the tech world or whatever, but it's so cool you ran into him in an airport, and he'd be like, you know, oh my god, like nobody ever recognized me here, but it's suddenly like Austin who walked up to me. So it's like a hyper niche thing. It's like the moment we step out of Computex, nobody knows who we are. Like the moment, you know.

SPEAKER_00 7:06

Yes, totally, totally. Um, in my neighborhood, you know, everyone just is like, oh, that's that dude that runs all the time. But other than that, people, you know, have no idea. And then you go to something like into Taiwan, to another country, and then all of a sudden people recognize me. And it's like, oh, this is weird.

SPEAKER_01 7:22

Yeah, so uh maybe this is a consequence of being in a highly populated country because in my neighborhood, somebody recognized me. I'm like, whoa, that's insane. I was just going to buy milk. Uh and then in the airport, somebody said, like, hey, are you the guy who writes the newsletter? And so I was like, Yeah, that's me. That's amazing. And so, you know, uh in in Computex too, we met so many people and it was it was great. But uh, you know, part of the Computex appeal is actually the exhibition, which is fantastic. You know, it has like multiple buildings and multiple floors of everything from AI racks to like chairs. Like I saw chairs, okay, like like like computer chairs that you could sit on and work on. I'm like, so, and all the gaming and all that in between. So I suppose that it's natural to have gaming chairs as well. It covers so much. And at one point I walked into the ground floor, uh, and there was like a ton of people, and I was like, yeah, everybody's looking at AI racks, but it's not so because there was actually Jensen who was at the other end of that looking at all the racks and stuff. And we I was with like a couple of people. I think you had gone back already, and I was like, uh, we were like, that's it, dude. We're not going into this hall right now. Because the way it is, like, like there's security walking on either side, and they're like making way for him, and he's like walking by. And um, I I I guess somebody told me like people were keeping track of whichever booth he stops in and like buying stock or something. I don't know, like each place he went to, like, was having like these mini pops, and uh there were maybe people were counting. Oh, he's like he stood at uh Delta booth for like 20 seconds. Must be something there. Uh-huh.

SPEAKER_00 9:03

It's yeah, just it's like momentary arbitrage, like you get a signal before everyone else. While someone else is posting the pitter to picture to X, you quick invest, and then someone everyone sees it. Oh, Jensen stopped there on X, and then they buy the stock. That's crazy.

SPEAKER_01 9:17

So talking about momentary events, right? Let's talk about uh the networking uh keynote I attended, which is like Marvel. And I was in the front row because they they let the media in. I guess I'm media now. Uh yeah, so I was sitting there and you know, uh I was pretty surprised because I didn't have any idea Jensen's gonna come trotting on stage. So he comes in and is like, hey everybody, what's up? And all everybody were like happy, yay, Jensen. You know, we get to see the Jensen. So I was pretty happy because I missed the GTC actually. I didn't, I like it'd pretty cool. I've never seen Jensen in person, okay? So I was like pretty stoked by it.

SPEAKER_00 9:51

Well, that was your first time seeing the leather jacket in the flesh. Yeah, cool, nice.

SPEAKER_01 9:56

I only write about all the stuff leather jacket creates on uh, you know, my Substack and all that, but I've never seen him in the flesh. Well, there you go. And and and how was it? I I don't know, it was cool, I guess. I mean, I don't think that that one time Jensen sighting is the feeling he's ever going to come back. Like, oh, that's Jensen, by the way. And then he comes in and says, like, uh, Marvell, everybody, the next one trillion dollar company. And I was like, ha ha ha, funny joke. You know why? Because just then uh Matt Murphy, the CEO of Marvell, had said, oh yeah, like we have already had three trillion dollar memory companies uh recently in the AI boom and all that. And I was like, yeah, yeah, that's pretty cool. And I didn't expect Jensen to come by and say, Oh yeah, next trillion dollar company. I thought he's just saying that to make Matt feel better. Like, oh, don't worry, you'll be a built trillion dollar company too someday. And then I I learned the next day, or I don't know, the same day, whenever the markets open, it's like Marvel is up 70%. And I was like, oh my god, there was my arbitrage event, and I was there taking Jensen leather jacket video the whole time. My video cost me thousands of dollars. I should have bought it's fine.

SPEAKER_00 11:08

That's hilarious. No, yeah, and they they were truly up like 30%. Maybe they got, I don't know, something crazy, like 50 billion in market cap or something just from Jensen's comment. Of course, or you could say it was because of what they presented and the merits were good and the story was good and whatever. Regardless, I thought to myself, oh man, you know, Jensen goes around and gives two billion to this company, two billion to this company. Hey, at this point, he doesn't have to give two billion, he just has to give compliments, and maybe that's even more valuable to the company than a $2 billion investment.

SPEAKER_01 11:38

Yeah, I know. That's that's more valuable. You can keep your mind, just say Jensen. Jensen says your company is good, and then you're better off than his two billion. So Jensen's happy he gets to keep his two billion, and the others are like seeing everything pop. Anyway, it's strange times we live in. Uh, I thought it was funny because I have never been in the middle of uh an event. You know, uh a hundred billion, I think it was hundred billion actually, just close to a hundred billion. Uh that the comment was, and it's like, wow, I saw a hundred billion sentence come by right by my face, and I just let it go by. Like I did absolutely nothing. Yeah, but the whole uh talk about Marvell CEO basically was like, oh yeah, optics is here, and uh you know the wall, the copper wall, has now moved out to the point that uh you can only do like on-package stuff, and everything within Rack is now going to be optics, you know. So he showed this cool slide uh about that. And so I think the whole keynote was basically about uh Marvell's uh prowess in optical interconnect, especially when they acquired INFI many years back. I think great acquisition. I think they have like a very strong optical portfolio now. They showed off their Teralynx switch, which is a 102.4 terabits per second switch, uh, which is I would say like what uh the Tomahawk 6 by Broadcom is similar. It's a similar 102.4 terabit per second switch. So that was their big announcement there. But yeah, they they showed uh a big uh big tray of CPO and he's pointing to it and saying, like, look at that, um that's how big it was. Yeah, the switches are yeah, those are big, right? I mean, if you have to do a pluggable uh implementation, those are really big. But then he showed the CPO, like which was like much smaller, right? Um it's like the size of your whole hand extended kind of uh size. So it's amazing the kind of compression and size CPO can bring about. And uh yeah, so the whole show was essentially uh very heavy you know, uh heavy optics. Like people loved CPO. Like everybody was every demonstration is CPO and this and CPO that, and the talk was all CPO. That's what a sense I got at least. So yeah, it was cool actually. The the whole keynote was nice. Uh I've I've never been to uh uh too many such high-profile talks, like right up front. All this is very new to me. But I had fun.

SPEAKER_00 14:13

Yeah, yeah, good. Yeah, I thought the Marvel keynote was good. I think Matt Murphy is a good public speaker. Um, I think the of the key of this like stories I've seen them tell over the last like say two years, I've always felt that they didn't Marvel has not done a good job of articulating who they are. Are they an ASIC company that also has interconnects? Are they an internet connect company that also makes custom ASICs? I think to in this story, they really try to position themselves as an interconnect company company. And and if I recall, they showed how like um their revenue was, you know, the majority of it was from interconnects, and then some of it was from AI ASICs. So I thought at least they narrowed in on like, hey, we're talking interconnects. We think of us as a connectivity company. Um, and then you know, they talked a lot about uh the copper wall and their portfolio. I did feel like they missed an opportunity to talk about how those two businesses can cross-sell each other. So, like, hey, think about it. If you're Marvell and you're making an XPU for someone, it's not just about the XPU, it's also about the XPU attach. And a lot of that is around interconnects. And so the whole point that I'm trying to make is they can say, like, we can work with hyperscale customers to make their XPU, XPU attach their data centers. It's really bespoke data centers, and that can help us sell into the interconnect there. Um, and you know, vice versa, as we sell interconnects for people, maybe that's opportunities to get our interconnects into their systems and then maybe eventually sell some AIA6. Maybe that's not a story they want to tell because maybe these uh companies that they make the data centers for don't always buy their uh like switches, for example. Maybe they're still buying Broadcom switches, so maybe that's not the true story. But I would love clarity. Are these two distinct businesses or do can they actually benefit each other and cross-sell? And they never seem to talk about that. So if anyone from Marvell is listening, I'd one, I'd love an answer. And two, if you have a good answer, then you should tell that part of your story.

SPEAKER_01 16:21

Yeah, absolutely. Um, I was listening to the Bank of America conference today where uh Matt Murphy also spoke about like the interconnectivity, their connectivity products. And then in the middle, he was also like, yeah, yeah, yeah. We also do some ASIC stuff too. So, you know, it's so I see I see what you mean by the disconnect there. But uh, I guess their essential business is data movement, however you want to frame it. I guess you could ASICs move data inside the chip. That could be one framing. We are a data movement company within the chip, as in the ASIC business, or uh between chips, as in the interconnect business, you know, or between data centers in their DSP business. They have a very strong DSP business, actually.

SPEAKER_00 17:03

And I guess it is a bit counterpositioning against Broadcom, who, of course, has an amazing uh data movement business, but spends a lot of the earnings call talking about the ASIC XPU business. Um, I think that's also just what analysts want to ask a lot about. Um, so it is we are in an interesting world, I think, both for Broadcom and for Marvell to as they have these two different businesses to figure out what is the cohesive narrative, basically.

SPEAKER_01 17:33

Yeah, that's true. Uh so you know that leads to the question of if all of the interconnects today, you know, uh optical is used in scale out. Uh, we know that. And uh there are still a lot of a lot of benefits to using copper or uh you know optics, it's not like one or the other yet. Uh but you know, as speeds go up, I think copper will run into its limitations. And um so optics is nice, they have uh their entire portfolio around it. But I think what is more interesting to talk about is the scale up scenario because when uh Matt Murphy drew that line and included the rack in the optics thing, it's very interesting that now uh internally we may be seeing the birth of optics for scale up. And recently there's been a lot of debate online about whether CPO is really here. Um when is go when is it going to be here? Uh can can we support uh optical engines and packaging them at scale? Because I don't think that we have done them at scale, like Spectrum X and uh Quantum X, and you know, Tomahawks from Broadcom have CPO, but they're just like one application and they're just coming into production this year. So nobody really knows if you can produce CPO at scale, if there's capacity to do so. Uh and also if there is like the question of yield, okay. Uh the reason I'm saying yield is uh you're packaging in the TSMC scoop process, you're putting uh the optical engine on top of the electrical engine. And so uh that I don't know how that can be done at scale. So I have no idea. They said something like, oh, we have 99% yield in in engineering samples, but that doesn't mean you can scale up to uh you know uh millions of these things. So those are all like unknowns at the moment, and also there's the question of testing you know optical engines, because especially when you have uh two-sided die, it's a problem now because you have to align the top of the wafer and the bottom of the wafer to test them simultaneously. Because you need the electrical side and the optical side to be aligned, and then you can test each one. And that alignment, from what I have read and spoken to people about takes time. So that is a big unsolved problem, or a problem that is currently being solved in terms of testers and things like that. How do you test CPO fast enough to make it a uh a reality? Right? And so uh CPO is is definitely coming because we see it we have to go to optical down the line. Uh there's no way around it. But uh yeah, uh there are a lot of challenges. So maybe a lot of the discussion online has been like, why don't we go to near-packaged optics for now? And uh we'll just not have to deal with the advanced packaging part of it at least. Uh, but then we can just do the silicon engines, uh optical engines in silicon photonics and maybe put it into a socket. So, you know, that's the kind of discussion that's kind of out there right now. Uh, and it's interesting to see how it'll evolve. I don't think I have all the answers at the moment, but it's interesting.

SPEAKER_00 20:49

Yeah, yeah. I I I totally agree with you. I mean, ultimately, near package optics takes a step in the right direction, gets that electrical copper trace much shorter, um, but doesn't jump all the way to CPO. Feels like a decent intermediate step, uh, especially when you think about manufacturing. Um, and of course, then there's XPO, as we've already talked about, which I think will continue to extend pluggables. Um, what else? Did you see much on NPO or XPO, or was it all CPO that you saw at the show?

SPEAKER_01 21:21

So I want to I want to talk about both of them. Actually, let me go with the XPO first, and because the CPO one is also uh nice. So the X is the first time I've seen an XPO module in person. Okay. And there's this whole video I took of me talking over it. Uh and we'll play it right after this. But it is essentially a honking big connector, like it's as big as my hand. You know what? Let's let's just look at the video and then you know it'll work out. All right. Check out this is the XPO module, which is the next generation of connector. Do you see how big? Uh This thing actually is like this is a honking big XPO module. Ridiculous. Oh my god. Look at this. This thing is liquid cooled, and it's the equivalent of eight OSFP plugs. It's insane how big this is. Wow. I'll show you all the cables after this for size comparison and all the other connector uh types, I guess. So you can see how big it is. Just for comparison, this is how big it is in my hand. Amazing. So the whole thing with XPO is that it it is effectively eight pluggable modules, OSFP modules in one. So it's a massive, massive connector that's liquid-cooled. It's it's quite amazing. See the number of cables that came out of it is this. So uh this is nice. I I like looking at it, but and I genuinely think that it has a shot at being an alternative to CPO if the ecosystem wants to stay pluggable. I don't think it's ruled out yet, so I I I'm glad I saw that part of the future of optical interconnects. But then the other thing I saw was the IR lab's uh scale up CPO demonstration with the Wii Win racks uh on the show floor. It was it was cool, you know. Uh in the picture of the rack, you can see uh uh all of these external laser modules that are like sitting in the front panel, like the faceplate, and then all of the uh optical fiber that's coming in from the co-packaged module inside the rack. So it was really nice to see something that it's not in production or anything, but it's like a concept-wise, as a as a demo-wise, it's really nice to see the fully scaled up optic CPO rack and be like, oh wow, this is what maybe data centers will be filled with in about maybe three to five years or some time frame. And so what was really cool about this was they had a video on this in the side that for people who didn't know what they're looking at, it explained the whole blowout you know image of the whole video. So, yeah, check this video out. So this is cool because watch this video and I'll show you what comes up after this. This is the IR lab's scale up optics. On the other side of the rack is the actual hardware. So you can see all the external laser modules plugging into this thing and those go into the rack. Yep. So that's a fully optical scale-up rack. So what I did was after that uh was once you know, once you see the insides of this rack and what is inside it, and how all the optical engines are sitting next to the silicon and stuff, they have their actual rack. So I took like a whole video of all the little details in the rack and uh you know how you can actually put they put give you handles, you can pull them out. So that's you know, yeah, I I'll just we should just play the video too, because there's a reason I took all of this stuff so we can put it in the podcast. All right, look at this is CPO rack level scale up. So each of these blue things are like external laser modules. And I believe this entire handle exists so that you can pull out the rack and change it out, like it does in that video. But you can see internal to it is the optical cable going to the CPO switch. And all of these yellow cables are actually optical fiber. That would be the switch, I believe. That would be the optical switch that uh hooks up all these XPUs in an all-to-all configuration. Other than that, um there was a micro LED demonstration by MediaTech, which was uh interesting, but nothing new. I don't know what I was trying to find out from them, like what is their reach uh and you know how fast they can run it and all that, but I didn't manage to get too much info out of that. But it definitely looks like it's uh it's still nascent, at least in the Media Tech demonstration. So I need to go to uh the next optics conference, uh either in uh Spain that's coming up, ECOC, or maybe OCP will have some some stuff about uh micro LED interconnects. So I definitely like to see more of that stuff.

SPEAKER_00 26:43

Interesting. I didn't know MediaTek had a micro LED offering.

SPEAKER_01 26:48

Yeah, they've been working on it actually. So uh they run uh you know, micro LEDs can only run one to four GBPS per lane. So the whole idea is that you have multiple lanes and you know transmit wide but slow. Uh but this array did like we counted it, it had only like like 100 LEDs uh something in that range. So in reality, what you see Avicenna uh demonstrate is more like 400 LEDs, uh so each running at maybe one one to two uh Gbps, and then you got like 400 or 800 Gbps out of the whole thing. So that's that's kind of what I was hoping to see, but it wasn't really a demo. The videos with the LED lights were just like coming on and off. So I'm like, I guess you're turning the lights on and off, and there was a microscope on there, and it's just like turning on and off, which is cool. I was like, yeah, I see some flashy lights, and those are like blue micro LEDs. Uh maybe they'll carry data one day. And so I was like, yeah, which is good.

SPEAKER_00 27:49

Sounds like um a children's science museum or something.

SPEAKER_01 27:52

Like, look, these blink. I don't know. I'll see blinky things every day. All of these ranks, racks were full of blinky things. They're like a cat, you know, you sure shine a laser at me. I'd be like, oh, that's optics, that's CPO.

SPEAKER_00 28:06

That's good. So, okay, we talked a lot about networking. What about power? Do you see anything? I remember seeing, like, I think right kind of right as we came into one of those conference buildings, wasn't there like a bunch, a wall of like power chips?

SPEAKER_01 28:20

Yeah, power was cool uh because everywhere on the show floor you see this stuff. There's like 800 volts this and 800 volts that, you know, everybody had 800 volts in their racks. Like it doesn't matter who was making the rack, every rack provider had an 800 volt system there. So I'm like, wait, are we are we already there now? Like this is the thing with shows like Computex. Like, you look at this stuff and you start to believe, oh yeah, this is what's in the data centers today. It's really not, it's like what's in the future. And uh it's nice to see, but it's also important to keep in mind that not everything is running at volume today. Like, you know, that's it's a nice thing. So apart from that, like like you say, when we walked and we saw this like entire wall of chips, and it had like these power converter chips like 800 volts to whatever, like you could 12 volts or 800 volts to 50 volts. Uh, all these chips from every provider was there. Like, you know, you could you could see all of them. And it was cool because you're like, wait, this is your competitive analysis right here. If somebody gives me a lab setup, let's start testing this right now and let's find the alpha from this wall. This wall has the alpha.

SPEAKER_00 29:29

I wonder what they would do if you started plugging in and trying to test. Yeah.

SPEAKER_01 29:35

I'll just get my oscilloscope and then just wheel in some power supplies and get to work, you know.

SPEAKER_00 29:41

Right, totally. I was surprised they had very different form factors and different shapes. I kind of thought, like, you know, when de uh when Jensen always does the demo of like, oh, here's everyone's hopper racks, and it's like rack, rack, rack, rack, rack, and they almost all look the same, just like different colors or something. But these um boards were different shapes. Some were like long rectangles, some were like L-shaped or whatever. It's very different.

SPEAKER_01 30:04

Yeah, yeah. So I I'm not sure how that slots into the actual application. They're very interesting. Yeah, it's a good, it's a good observation. I don't I don't know why they're all like so different, but uh these are like big chips. And I went to the Lighton uh power booth after that, and they had this like again, the whole array of chips that they have. In the beginning, they had this like massive um uh chip, and it it basically converts uh um from the medium voltage grid down to 800 volts. So this is like the other side of 800 volts, you know, the the grid side of 800 volts. And um yeah, so I asked the guy, like, what is this made of? And he's like, Yeah, it's silicon carbide. And uh so then I asked him, what does all these other chips do? Like from 800 volts down. He said, Oh, those are all like silicon. So like, I'm like, they're not silicon carbide? I'm like, no, no, silicon. And so what about GAN? Like, he's like, nah, GAN is still new. We didn't, it's just silicon. I'm like, oh, it's cool.

SPEAKER_00 31:02

Wow, okay, so grid to 800 volts with silicon carbide, but then 800 volts all the way down was just silicon.

SPEAKER_01 31:10

Yeah, that's why I took a whole video of this thing. So you should definitely watch that too. This is Lyton's chip conversion uh portfolio is really cool. So these are silicon carbide chips that uh use transformers and have single stage conversion down to 48 volts. All of these other conversions from 48 volts, whether it is to 12 volts or 6 volts direct, are all silicon-based. So the argument of always whether it's gamma or silicon carbide, isn't it? All of this uh switch stuff you're looking at here that converts voltages down uh is all silicon. So don't discount silicon yet.

SPEAKER_00 31:51

Man, it's so cool. There's so much to learn from just talking to people on the show floor. I think that's the best part, is just asking what feels like dumb questions, you know, but I think there's just so much that you can learn. It's you know, of course, it's like asking Chat GPT that's patient and will answer the dumb questions, except it's the actual people that work for the companies and have the correct information, not something from 2024.

SPEAKER_01 32:15

Yeah, it's I wish I could stay there and everybody could explain to me for the whole week, but that's not to be the case. But yeah, I was like a kid in the candy store, so I walked up to this demonstration and there was this power whip, which is just this honking large cable. And I took a picture of me holding it, and you look at that thing and go, oh my god, like what sort of a cable is this? Look at the size of that thing. No wonder people want to go to 800 volts so you can carry lesser current and you can reduce the size of this honking massive cable. It's ridiculous. Totally.

SPEAKER_00 32:46

Yes, yeah, literally like the size, the size of my forearm or something, you know?

SPEAKER_01 32:50

That's how big it is. Yeah, yeah, yeah. So it's crazy. And yeah, so the other thing was on the other end of it, it was insane. Like you go to the Vertif booth. Um, by the way, before I even explain this, check out this video and I'll tell you what it is. By far the vertive booth is insane. Inside that thing is the entire cooling solutions they have built up. By the way, those are not real, those are just like pictures. I wish they were real though. But uh, yeah, I'm gonna go around there and see if I can uh show you the cooling hole, I guess. There you go. Wow. Shiny colors. Okay, so what that was was basically what they call a prefabricated AI infrastructure block. Okay. Uh so what it is is it is a container home, like a container, you know, modular home built for AI racks. So the way it works is you wheel in all your racks into the data center. They could be like 800 volt racks or whatever, right? And but how do you hook this thing up? Like you have to give it power, you have to give it cooling and all of this stuff. So what Vertiv, and by the way, this is not just Vertiv, Delta has this, Schneider, Electric, Eaton, all of them have the solution. But I I just saw the Vertive one in a little bit more detail. You just roll this thing in, and it is pre-fitted with all the cooling infrastructure and the power infrastructure, all this stuff. All you do is like you roll it in over the racks and you hook it up and it's ready to go. And the guy, uh, I think this was a delta, uh, he was telling me that uh this this kind of reduces the time of deployment by half. You know, it takes half the time to hook up power and uh cooling solutions. So these are like prefabricated uh infrastructure solutions. Like it's insane, right? What do you have to think of? Like you just can't think about the chips. On one side, you see all these chips that convert, you know, you see even little chips that do the voltage regulator modules. But then on the other side that they're wheeling in these uh, you know, room-sized infrastructure just to power up these racks. It's fascinating.

SPEAKER_00 35:05

That is fascinating. You know, it it's amazing. I hear lots of different companies as a value prop state that they're trying to reduce the time to stand up a data center. You could call it time to first token, if you will, but in a different context, you know, time to the very first token that's ever generated. And of course it makes sense, right? If you're um, you know, Elon and you are building your XAI data center and you've bought you know billions of dollars of GPUs, it they're just sitting there depreciating every day. So obviously you want to turn them on as quick as possible. So it's fascinating to think about even from like the physical infrastructure, power and water, like how can you modularize that in such a way that it just makes it very quick to roll a thing in, plug everything in, off you go.

SPEAKER_01 35:58

Yeah, that's very interesting. And it's just that these cooling solutions are not only at the data center scale. Like I saw one solution there. Um, I'm not sure if you were in uh when you were there when we were looking at this. Uh like there were like a bunch of people, but probably not. So I'll tell you what it is. There's this uh edge AI deployment, right? Like, so let's say you don't want to deploy this whole data center thing and you just want to put, I don't know, like eight or sixteen in uh your you know office server closet and you know provide some inferencing tokens to your employees. That kind of thing is going to become more and more popular uh for just overall cost optimization. So they had this one company called Ketaflow Technologies, which is basically like you know, a chest freezer in your garage, right? So you open up the chest freezer, and what you see inside there is liquid. It's just like a transparent liquid. It looks like water, but it's kind of like an oily thing. They let us touch it even. So I went and like stuck my finger in it, and it's like all oily. And they know that people are gonna do this, so they get handed out tissues after, so like we all wiping our hands over there, you know. So it's funny. So, but yeah, there's this liquid. It wasn't warm to touch or anything, but inside that liquid is basically immersed an entire rack of GPUs and power and cables. So it's just like dipped in there, and it's doing liquid cooling, right? The whole thing is dipped. And so we were like trying to ask, like, how does electronics still work? And he said it's a non-conductive liquid. This is not like water, you know? Yes, so it's a non-conductive liquid, and um yeah, I think it's gonna happen to it, so you can put like racks in here. He's like, Yeah, it's running as we speak right now. It's uh it's it's nice.

SPEAKER_00 37:39

This is this is really fun. Yeah, it is it is mind-boggling the first time you think about it. You're like, wait a minute, I just always think liquid, short circuit, conductive, and then you're like, wait, there's non-conductive liquids, huh? Fascinating. Yeah, you could just drop the whole thing in and it works. And then of course it dissipates heat because the liquid helps carry that heat away. Exactly. Amazing, amazing. So let's see. So we talked networking, we talked power. Um, let's talk Intel a little bit. I went to Intel's keynote and uh had a front row seat for Lit Bhutan, and you know, the the audience loved him. And you know, he started by speaking in Mandarin, and I thought that was really cool. Um, like talking to the local, you know, uh, we were obviously in Asia. Um and you know, uh you can watch the keynote online. I just thought really quick, since we're kind of getting close to the end of time, I'd talk. Um, one of the things that really stood out to me was one of the big introductions was the Intel Xeon 6 Plus with e-cores, aka Clearwater Forest. And so, you know, high level, 288 e-cores, 12-channel DDR, 5,000 to 8,000 mega transfers, a big L3 cache, you know, the kind of things that you expect Intel products to be shipping. Um, this is very much like the space of going after AMD and turn dense, trying to catch up there. Um, so so good specs. Um, but I just couldn't stop thinking about like what a positive signal this is for Intel Foundry. And as they uh went through the chip, um, some of the specs that I thought I'd I'd talk about how it was built, which is really cool, that just shows that Intel Foundry is like putting in good reps and these the types of technologies that are being you know perfected on their internal customer, if you will. Although I know that Intel Foundry doesn't position them as an internal customer anymore. They very much want to say, like, we we treat Intel Products as any other customer. You know, we're trying to become a real Foundry and not give them preferential treatment. So let's just call it like their best customer, their biggest customer. Um so Intel Products came and said, hey, we want to build um uh the this data center CPU, and we want to use, I think there's something like 17 chiplets, and it's um two IO tiles or dies that are built using Intel 7 technology, three active base tiles on Intel 3, and so this is for like cache, mesh, fabric, stuff like that. Um, and then and then there's 12 compute tiles on Intel 18A, and these are the darkmont e-cores. And so then the cool thing is the um the active, and there there's like a nice picture of it. We can try to include that. Um, but the active base tiles on Intel 3, those are connected to the I.O. tiles using EMIB. So um, and which by the way, I think we're gonna try to talk about EMID in a future podcast. So here's a teaser for listeners. Um, I wrote about it recently. But so so obviously Intel's got these chips that are built on different process technologies, all Intel Foundry. They're connected with EMIB. Um, and then on top of it, they're using Foveros Direct 3D to stack the compute tiles on top of those active base tiles. So this is like this is hybrid bonding, which everyone likes to talk about, like bumpless copper to copper bonding, but it's also silicon on silicon, like like logic on logic. So active silicon stacked on top of active silicon. And that would be different than active silicon stacked on a passive silicon interposer, for example, or even um cache memories stacked on active silicon. This is active silicon on active silicon, um, which has its own thermal things that you have to think through and power delivery. And so then, of course, speaking of power delivery, um, Intel 18A is the first generation for Intel Foundry of using backside power delivery. Um they they brand it power via. And then they're also from a transistor perspective, um, this is using gate all-around transistors or ribbon fet, I believe is what they call it. Um, so as I was as we're going through these slides, and it's really talking about pitching things from an Intel product perspective, all I could think the whole time about was Intel Foundry and how there's a lot of really interesting technology going on here. And I know everyone just likes to be like, what's 18A's yield? It's bad, it's good, you know, whatever. But I'm like, no man, it's so much more than that. Like, there's even an economic story here about like being able to build a chip. And of course, this is the benefit of um disaggregated design or chiplets, different than monolithic um design, which is like you can use the right process node that has the right economics. So, for example, I think that they're even just reusing the I.O. tiles from Granite Rapids. Like, I don't think those are even new designs. Um, and then, you know, S RAM doesn't scale that great. So just leave it on Intel 3, that's fine. You know, only use Intel 18A for compute. And then obviously, from a yield perspective, you've got little dies instead of a big monolithic die. And um, yet it also goes to show how good Intel Foundry must be at advanced packaging to be able to take all these dyes, stitch them together, not only with eMib, which is not new, right? They've been doing EMIB for a while. Um, but then especially with this Fovoros 3D logic on logic stacking. Um, I just thought like, and and and it wasn't really even hammered home because that wasn't the purpose of the talk. But I guess I'm doing Intel a favor here, and just like as sort of a foundry person with foundry interests, I thought, like, wow, this is actually a really compelling Intel Foundry story. So that was my takeaway from the whole Computext, which was not the story that Intel was probably even trying to tell, but that was my story, which was just like, dude, this is sweet, sweet technology that they're building. That of course, and I'm sure that their customers, Intel Foundry customers, should be able to see that and think, like, oh yeah, this is the type of technology we want to use. And by the way, if it's on Fortune eventually, that'll be second generation ribbon fet, second generation backside power. So like Intel's, Intel Foundry is learning all of this on behalf of their biggest customer Intel products, but they're also like improving on it too.

SPEAKER_01 44:07

So Yeah. So what I'm hearing you say is that it's not so much as a whole CPU flex or anything. It's just like, oh, look at my advanced packaging that I'm capable of doing. You know, it's a packaging flex, really. It's uh when you can package up all this stuff nicely and do it at scale or whatever. Yeah, it's it's good. So good to see that eMib is coming up the way it is. Um, there has been news that Google is also engaging with Intel Foundry for their TPUs. I don't exactly know in what context, uh, whether it they're actually going to use 18A, which would be really, really cool, or they're just going to use the e-mib packaging part of it, which is also amazing for uh their TPU products. But it's so good that there is a uh product that is not co-ops from TSMC because it's so supply constrained. Uh, you know, we can, I'm not just talking about this from a business perspective. Oh, yay, like, you know, we have Intel winning, uh, we've had enough of TSMC. It's not that. It's just more that I would like to see technology accelerate only for the reason that I don't know what other cool applications are there for AI. And I would hate to see it constrained by the inability to package something or by the inability of lasers or something like benign like that. I'm like, it's a simple solution, it's not a technological problem. Why don't you just make more is the thing I want to say, but obviously reality is more complicated than that. You just can't make more, okay? That everything takes time and capacity build-outs are a real thing. Yeah, so I'm happy in that sense to see that there's a second packaging option, so we can just make more chips in some regard.

SPEAKER_00 45:48

Totally, totally. And you know, to your point, TSMC is an awesome company and it was so cool to be in Taiwan. I wish I would have been able to go to TSMC. I would love to meet the folks from TSMC and take a tour someday. Um, but but to your point, like TSMC is doing everything they can within reason to increase their capacity, but ultimately there is a limitation on just the amount of chips that can be built, on the amount of interposers that can be built, on the amount of packaging that can happen. And it's not to any fault of TSMC because they ought, it is very, very expensive and it's a huge investment to spin up fabs, and you have to be very confident that the demand will still be there by the time you ramp these things up in two years. And so it is not fiscally responsible to just build tons and tons of fabs because Jensen wants it today or Sam Altman wants it today, or who or Elon wants it today, right? So they are doing everything correct and by the book. But zooming out to your point, if there's really only like one big supplier, then the whole industry is limited, which means there are innovations that can't come to bear or can't come to bear as quickly. And so, just from an industry perspective, the more packaging capacity, because of course even OSATs will be able to do eMib if they want, but um the more packaging that Intel Foundry can show off, the more waivers that Intel Foundry can build, just the better we will be as a society because we'll have more capacity and we can bring innovations to light quicker and potentially cheaper if you have competition and you can kind of keep prices in check and so on. So I'll stop my spiel there. We could talk a whole episode on just that, but I think it's net positive for everyone to see Intel Foundry having success. Agreed. Well, let's call it there. Uh, everyone, I hope you liked this episode, our Computex breakdown. Um, enjoy the videos, enjoy the pictures. Thanks to everyone who watches this on YouTube. I know I call you out, but we get, I think our last uh video, you know, was our best one yet, as far as engagement. So thank you for all the comments on there. Um please share it with a friend. Uh, follow us on Spotify if you or wherever you if you like to listen, check us out on X, whatever, send us comments, we read them all. So thank you, everyone, and we'll see you next time.