18:44
Travel expenses for Computex 2026 have been covered by Sonology. Thanks, Synology. Intel Copyex. I was expecting the announcement to be under embargo, but it wasn't, which caught me by surprise. So, you know, it's taken me a few days to put my thoughts together. The Intel Arc G series processors set a new standard for gaming. According to Intel, I got to go hands-on. Uh, I spent the most time with the MSI Claw8 EX Panther Lake. It's a new Panther Lake CPU. I also got to talk technical details with Intel's Tom Peterson. And so 8 hours to 11 hours with CSGO in a handheld. Yes. Okay. I've set your expectations just a little high. that is possible with CSGO, but the magic there really comes from doing
a lot of really heavy lifting in software and doing as little actual computation as possible, which is actually pretty good. You see, Intel is doing a lot of hard work here with the graphics and processor aspect, juggling power and making sure that the frame pacing is on point. If you have a relatively low frame rate, but the frame pacing is perfect, you will have a good gaming experience. And that is what Intel spent the majority of its time with us trying to convince us of, including Steve from Gamers Nexus. And so you should check out his video because I was lurking in the room. See, I had a lot of developer questions, developer oriented questions like, "Okay, Intel, what sort of recommendations are you making for the game developers in terms of minimum
frame rate? Like what's in the SDK? What's in the documentation?" And unofficially, 25 FPS. Like before frame generation even enters the equation, 25 fps is sort of the minimum that you want to target, which is an entirely reasonable answer. For whatever reason, we ended up playing Lego Batman the most, and it was a delightful experience. And the battery life was really good. On a game like that, you're not going to get 11 hours or 8 hours or whatever, but four or 5 hours is entirely reasonable. And there were also a bunch of other form factors that Intel showed off from other partners other than MSI. So, this handheld looks like it's going to be pretty amazing. Now, I had a lot of questions about Linux. How
are we going to do this with Linux? Um, Linux will probably be possible and I'll probably be one of the first to do it just as soon as I can get my hands on a device. Intel doesn't have the same plumbing uh for frame pacing and power management that Linux has. So, hopefully Valve is on top of that or Valve folks are on top of that or it's going to be me and the community that are going to be on top of that. First stop, Computex 2026, Intel and their handhelds. Panther Lake, two PC cores, and a bunch of efficiency cores. Mainly, I got to look at the MSI Claw 8 EX AI Plus. This is a newer version of the MSI Claw 8. The EX means a lot here
because it's Panther Lake. Like, Panther Lake is legit and is very nice to see so many Panther Lake based designs. Intel has done a good job with this CPU as we reported in our other coverage, but it was nice to be able to go hands-on with devices and to get a chance to chat with Tom Peterson at Intel who is an engineer that knows what he's doing and commands a lot of respect because he tells it like it is and he does know what he's doing. Now, all of these handhelds, they're running Linux and the battery life could vary from, you know, an hour to 11 hours. And how are they getting 11 hours? Well, that was Team Fortress 2, and that was making it run at about 30 FPS, give or take. But the engineering
here is real, and one of the things that Intel is leaning on is their XESS graphics stack on the software and the hardware side in order to make this happen. And yes, that does include frame gen and multi-frame gen. For me, I can't wait to install Linux on it and see how it goes. And I know that I'm going to take a performance hit and it's probably going to make the battery life worse, but don't care. They were also showing off the Acer Predator Atlas 8 and the 1X Player 3. Several different configurations based around this 2P core Panther Lake part. But one of the really surprising things that they said was that a lot of games will run fine on the efficiency cores cuz you get the regular efficiency cores and low power
efficiency cores. And the regular efficiency cores have access to the L3. So the cache topology here is maybe a little different than you expect if you were looking at, you know, the previous generation hardware from Intel and how they had arranged their P and E cores. This is a much more intelligent setup on this generation of devices. I have to say when I first saw the specifications for these on paper, I was a little concerned because we only got four lanes of Gen 5 and eight lanes of Gen 4. Intel has cut everything here. I mean, it's Thunderbolt 4. There's only the two display engines. This is meant to be as low power as possible. And Intel is targeting a 12 to 15 watt power envelope
configuration in order to get the battery life and the performance that they want. It's good to know. And I also think that for what Intel is trying to do with handheld gaming that upscaling and generation frame generation technologies certainly makes sense from a power conservation efficiency standpoint. But the question is what will it do to the actual gameplay experience? Time will tell. Now, Intel had a lot of other announcements at Comp Computex and I got to do uh a bunch of interviews to do with Xeon and power management and all of that. And Intel, you know, we talked about the all ecores Xeons a couple of months ago at Intel's special event. Actually, I think
that was late last year or maybe early in 2026. And it officially launches as Intel Xeon 6 Plus. Like the handheld, the magic here is the shared L3 cache because we got two performance cores and four efficiency cores and four low power cores in handhelds. Well, on the Intel side on server, the shared L3 cache is the performance breakthrough and the shared L3 cache is the performance breakthrough on handheld as well. Zeon 6 Plus is socket compatible with uh Zeon 6, both our ecore and PC cores, which uh makes the upgrade simple uh streamlined. But yeah, we packed a whole lot more functionality and capability in this product up to 288 of our new next
generation ecores. Also, what's really exciting is it's on our latest 18A process node. And so that uh drives a lot of performance, power, efficiency as well. When we started the product, the markets evolve and you know, I would it would be disingenous for me to say I was the only one 5 years ago that saw a genenic AI coming now, right? The only one in the whole world. Having said that, what's really driving the resurgence of the CPU with the genic um workloads is there are some principles that have been true for a long time that I think continue to be true about next generation compute capabilities, requirements.
Um we will always that the world will always want more compute, more performant compute. But we've also introduced for the first time a 3D packaging we call fauos direct 3D which has allowed us to pull forward um our latest 18A process node performance and power efficiency. This is our first Xeon product on 18A and we do that by leveraging this new 3D packaging technology that I think is going to really take over the industry as we continue to pack more compute more power efficiency into the same socket. So bringing all of those things together, building an architecture that accommodates all those technologies, bringing those technologies to bear to deliver real world performance with our next generation Xeon, and then all the
challenges that come around how to put that into a single product and a single package. Those are really the cool engineering challenges that I think most of us get excited about. Intel also announced their E385 networking platform, which I also got to see and do a little bit of hands-on testing that they had in the lab. So here we are at the showcasing our um physical AI booth, right? So basically what physical AI is trying to teach an AI model how to operate a robot. So these are so-called PLA models, visual language action model. So what we do, we train them with uh joint positions, camera input and uh text input. So on the right side you see the data
collection part. Data you always need a data set before you can train an AI model. So how we do that is using teley operation. So I can control the uh this arm using the uh the leader and while I'm doing that it's recording joint states and the camera feeds and uh I showcase basically the task that I put in into the prompt. So for example I could say put bracket from A to B, right? We do that 50 times. We do that in the so-called physical AI studio product that we uh that we built over here. And then we could start training the model. For fine- tuning the model, we use an B70. So there's a machine over here. Um and after it is trained, we have a model. So for inference and deployment, we are using the uh core
ultra series 3 which is code name pendant lake. Uh so there is a nook over here which we use for that. And uh for this model, it's a Pi0.5 model. Um we are running the model in the B390 iGPU. Um so despite being at quite a large model, we're able to uh to meet realtime inference basically. So that's what you're seeing here. So this is a bmanual demo. Um it tries to assemble um uh this bracket onto that uh engine unit basically. That's what you're seeing. Yeah, it's doing fine. We're still waiting for the chat G C G CPT moment in physical AI. So here and there makes mistakes. But on the other hand also this has only been trained with a few hours of uh of episodes. Um whereas the bigger models typically work
with hundreds of hours of data, right? So um but that's not the point. We are not selling a model here. We're obviously trying to uh to showcase how well the models perform on uh on the new Core series 3 processor. All right, I'm here with Craig and you've got agents running on Wildcat Lake, but the AI is not running there. It's running on a 4x B70 system like I have in my lab. So, show me what I can expect. Okay. Well, let's go ahead and take a look at it. Well, first thing, let's go ahead and introduce you to SuperClaw as an openclaw derivative. It's obviously we have it completely quantized and it's optimized for our hardware as well. So
the first thing that we're going to go ahead and kick off is we have this on the Wildcat Lake and I'm going to do a comparison which is compare superclaw which we're showing today which is Intel's own derivative and let's compare it to OpenClaw. So it's gone and it's gone ahead and done that comparison for us and it's cranking away right now. But the model that we're using is Gwen 3 coder next. This is 80 billion parameter model that we're doing and it's all running on this Grand Rapids workstation. H four. Yeah. We're having on four. And then for you can see that we're crunching all of these. It's actually doing a huge web search with individual agents. And then over here, well, nobody
knows about Superclaw yet because we just talked about it at this show. So I have a product brief for it, but it's Intel confidential. So, I'm going to go ahead and include it in our deep local research, but it's not compared on the web. It's actually compared on the B70 workspace. So, the web searches are past historical context mixed with local context that was nowhere else. That is correct. And so, you process that instead of doing it in the cloud where you're making the comparison. No, that's that's pretty sensitive information. we're going to go ahead and process it on these B70s to make sure the if sensitive bits never leave the machine and go into the cloud. So in
those cases it keeps it very private, very secure and then we can have our entire analysis of superclaw versus openclaw whether it's architecture or those cases and then of course Gwen 3 coder next is pretty nice to always have access to wherever you want to. So, as far as running to API zero and having an amazing experience with all those agentic kind of mixed expert models that we really want to use in a lot of these cases, this I'm excited for your setup. I wish I had 4 B70s outside of the demo lab. That's pretty impressive. Yes, 128 gigs of VRAM is kind of becoming the sweet spot for local AI.
It's kind of not. More is better as always, but also more costs more. So, in a lot of these cases, how do we make our hardware really be able to ring out kind of every ounce of that AI goodness that we want? And at the same time, please, tokens are expensive. Getting used to the million token context window for local and it's spoiling me. Well, I mean, it's going to become the default, right? We've seen it. People love it. They want it. They've launched it in a couple of cloud services and it costs us through the roof. So how do I make that as a constant available model to us constantly? Something much more attractive. Yeah.
Yeah. This is great. And so we can have it local. In this case, it would be designed to be behind your firewall. So instead of having I'm opening up co-pilot where it's I just opened you up. We have the same experience that we do that all the other employees. Why not virtualize this inside your enterprise? So you can have oh as an employee I get my own little Jarvis and I we talked I remember what we talked about yesterday. we're in the middle of a coding project, let's go ahead and finish that versus starting the entire conversation from scratch. So that's the real value that you're getting here in a lot of these cases. And you know, this is funny as OpenClaw becomes that consumer technology now a lot of the value that
we're seeing in completely autonomous agents is commercial. So let's bring it that way and make sure that we have the security and the privacy on top that really makes it valuable for that use case. I got fun times with local AI. Good luck with your B70s, man. That's going to be a good one. Poi ASAC. This is on GitHub. It's open source. You can load your vector database directly from SSD. It doesn't consume memory. So, this will save VRAM. This will save KV cache. This will save your AI solution because you're just pulling the data directly from high-speed SSDs.
This is going to put your Gen 5 SSDs in for a workout. And it's nice to see that it's on GitHub. It's open source. It can be extended. This is Kyokia's storage next platform or their implementation of Nvidia's storage next platform. It's running on a GP series high IOPS SSDs. 44 E1 bays in our chassis. Supports up to two H200s or RTX 6000s for AI applications, but uh that's a lot of IOPS for storage. Next, this is the kind of chassis that is going to store data that is really aimed at like aentic AI or workloads where you need an absurdly high amount of IOPS. Also, I would call attention to client SSDs there. The I mean handhelds and ultra low power, you know, laptop manufacturers have noticed,
handheld computer manufacturers have noticed the power consumption. And so, Kyoki has a new spin of BG7. It's like 4.7 watts give or take, but you still get capacities in the terabytes range. So, low power, high capacity, high performance. To give you an idea of the state-of-the-art, Kioxia brought a wafer, an entire wafer of flash. It's 32 layers of 2 terabs of flash. So, 32 * 2 tab, we've got 8 tab of usable capacity per tile, per chip. and also look for upcoming coverage with Synology. I wanted to go visit Synology and get to the bottom line because they're great in a business or enterprise context, but I'm not sure I like where they're going with some of the other stuff that they're doing with their products. So, I wanted to ask
questions and see what they're up to, why the change, what's going on. So look for that video.
