Linux Weekly 6: 9950X3D2, Local AI on Intel GPUs, and Forbidden Routers

Good news everyone. It's time for another Linux weekly. And in this week we have more hardware launches. It's good tea. It's not good tea unless you can really feel it in your kidneys. And when the tea is that opaque, you can really feel it in your kidneys. I got to make some weaker tea for next time. All right. 950 X3D2 has launched.

I did a full review of it. I'm still working on the Linux aspect of the review, but I wanted to share this is the ASRock rack. It's a half-depth B650 thing, I think. Uh, yeah. So, systems like this were not necessarily designed to deliver the 200 plus watts of power that the dual 3D vcash chip can deliver. See, you got two chiplets that have 3D vcache and gamers mostly don't benefit for that except oddly Borderlands 4, but code compilation server task hosting building a little build server. I've got this thing configured as like a quasi storage server with the B50, the Intel B50, the ARC B50 because Sri is working correctly, which we'll talk a little bit more about that. There's actually some kernel 7 Sriov news. Dual 3D vcache is awesome for Linux and awesome for consistent scheduling and off awesome for having 16 of the same core. So you have core with the same cache configuration. That is very nice. A lot of cool things fall out of that. It's also uh for like scientific compute if you don't want to shell out for a thread ripper can be pretty awesome. And even if you do want to shell out for a Thread Ripper, for some really um niche use cases around computational fluid dynamics, $1,000 is a really good deal. So for game hosters and people that would want to run it in a system like this, $1,000 is a bargain. For everybody else, you could probably get one of the other CPUs from AMD and save a bunch of money like the 9950X, which is like $500. Like that's a steal on a 16 core processor. And part of that, I think the reason that it's a steal is a little bit of pressure from Intel because the al the 270K Plus has a lot of cores. It's a mix of P& cores. You're back to the whole scheduling thing. Like there's a whole other dimension with that. But both of those CPUs, Intel and AMD benefit tremendously from Colonel 7. And so Colonel 7 has

launched on Cachio OS. You can try that. And it's not just Colonel 7. There's also some stuff that's been backported. I mentioned that last week. I now have had a chance to try it and it is generally pretty awesome. The context where I've tried it the most is on PVE. Now, Proximox backup server has some sort of catastrophic bug with Colonel 7. You probably don't want to do that, but for updates for Intel Arc, B50 and B60 primarily, B70 is a little bit of an odd duck. Uh there's some threads on the forum for that. And the really amazing thing with the Intel Arc GPUs is not the gaming use case. It is single route IO virtualization without a subscription and kernel 7 and the XE driver. You no longer need the out of tree i915 driver. Uh if all of that sounds like Greek, there is excellent resources probably I think the best resources on the internet on the level one forum thanks to amazing people that have helped me put stuff together and some of my own write up that people have taken and extended and documented. So thank you for that. And it's not been an easy path. Intel has not made it easy with firmware updates and everything else. We still want the B50 to have four VDEVs. It does get squirly when it has less than 4 gigs of VRAM. I will give you that. But two Vdevs with 6 gigs of VRAM doesn't solve any problems that four devices with 4 gigs of VRAM uh don't. So, I feel like that we should get four SROV devices on the B50 and nobody will talk to me about why we don't have that anymore because we did have that with the previous version of the firmware. So, looks like the B60 is probably the better choice and that's what I'm running on the Xeon 658X. The review for that 658X Zeon is coming out this week, a little bit later, so stay tuned for that. Uh, it is mostly an amazing Linux part. Windows eh I mean you can but Linux really amazing Intel Xeon 658 X 24 core 48 thread all PC cores all the time uh off the scale power efficiency about 5% faster per core so it's not oh my gosh Intel has won the core game but it is a very strong showing from Intel especially memory bandwidth especially 24 cores being able to saturate the platform in terms of both memory bandwidth and IO which is something AMD can't touch with their 24 core pro CPU. Yeah, the 24 core pro CPU leaves a lot of memory bandwidth on the table when we're talking eight memory channels. Just to be explicit about that. So, look for that review. It's going to be a lot of fun. Other

hardware and launch news. Framework 13 Pro is launching and it looks pretty awesome. And they got me an invite out to California. It's just that where I am, it's very hard to get out to California. But fortunately, I knew somebody that was in town that could uh gather all of the resources and intel for me, which is uh pun intended. So, Framework Pro, you should check that out. New battery life, new design, new options for the case, and also some upgrades for the Framework 16, which I think has been more popular than they expected. So, uh, one of the big upgrades for the Framework 16 is they now have an Oculink connection dev kit. And that is an eight lane Oculink dev kit. I literally said that in the review, in the day one review. You could have asked me when you were doing product development and I would have told you the same thing then. So, uh, you know, I can keep a secret. It's fine. We can talk product development and tell you what all is there. But an eight lane Oculink in a laptop, that seems like a pretty amazing docking station. Now, it's not hot plug or anything like that. But having eight lanes that you can connect your G your framework 16 to, pretty

awesome stuff. Now, I will just do a quick mention. We have some awesome threads on the forum for people setting up local automation including AI and that is around B60 and B70 but also R9700 GPUs. So nice job Long Barn and a bunch of you other folks that are working on that kind of stuff. I am a firm believer in local AI and all of the stuff that will enable. And uh a lot of people misunderstand what I say because I'm excited about things. But that's a story for another day. I really am positive on AI because I think that even the worst people at the helm can't foul it into something that is ultimately a negative outcome for society. Yes, this it's going to be a scary time. And yes, the people that are in charge of our governments are um too stupid to give us a controlled ascension, but that is our own fault for not being more involved in government and policy. We are smart enough to have a controlled ascension that is not um anti-soccietal progress. I believe that we as a society are capable of that fundamentally. It might take some extreme dysfunction, some more extreme dysfunction than we've already endured to get there. But ultimately, if that's what happens, and that gets us there to a more reasonable and just society. Okay. And that kind of

dubtales into This Week in Projects because oh my gosh, this weekend projects, uh, I will concede that there is an aspect that maybe the LLM doomers were right. Like LLMs are never going to get us to true intelligence. They sure are useful as a tool, but I definitely would not abdicate them to decision-m as I have said repeatedly and constantly. But there was this amazing post um uh lun lun he left ma and he's like loll are a dead end. We're going to do ja or ja and there's a paper on that and it looks like he's right. It's very believable but also that will if that is correct and development happens there is a path to very quick alleviation of the pressure on memory and everything else. So there is hope that if that is useful for robotics as useful as it appears and scaling and everything else then we don't have to scale LLM to the moon in order to do basic tasks. It's like, oh yes, I need a trillion parameter thing running on my wristwatch in order to do X. Kind of like how, you know, we didn't have anything better for memory other than Mercury delay lines. And if you look at the amazing stuff of computers, it's like, oh, we don't have enough mercury on planet Earth to scale up to the amount of memory that we need. Yeah, going to need better technology. But the approach, the science, the motivation, the agency to build that was there. I know it seems like I spend every waking moment hyping AI, but almost all of my time this week has been toward turned toward janitorial tasks. And I want to hear from you if you are experiencing MediaTek jank. So the HPG1A and some of the other AMD based systems, the Stricks Halo, you know, again, my in my infinite time working on AI stuff, of course, I'm being facicious with that. Um, a lot of those AMD platforms have MediaTek Wi-Fi. And I've experienced the MediaTek Wi-Fi being amazing, but it constantly breaks every time there's a software update. So, I want to hear from you if you are experiencing MediaTek jank. Um, we should chat about that on the forum and try to organize a little bit because I think I've got a path to making it a little less jank. And some of this is upstreamed MediaTek stuff, but I think some of this is just not really doing um good unit and integration testing. Like I have a feeling whoever's working on the MediaTek driver just has like a test platform and they're not really doing anything else with it other than working on the test platform because some of the bugs that I've encountered have been really dumb that I've been able to fix myself. Like the a newer version of the driver that is faster also assumes that you've got 40 MHz channel width and it just doesn't work if you have a 20 MHz channel width. So, if you have MediaTek and you're experiencing awfulness, let me know. The awfulness also seems to spill over onto the Windows side of things. I have no ability to fix things on the Windows side of the world. I'm so sorry. But I think we can fix it on the Linux side of things because GPD Win MediaTek, HPG1A MediaTek, Minis Forum Stricks Halo based machine, which I love and is going to be an amazing like home assistant all-in-one appliance. Uh, also R9700 logs. There's some folks on the forum that are doing some fun builds with R9700 for various reasons, one or two GPUs. I want to hear from you. So, if you've picked up some R9700s and you're building something, I want to know what you're working on. I want to see pictures. It's like, show me the cool stuff that you're working on, I want to hear about your projects because this is how we can learn and grow as a community. And I really do think that uh kind of like how the open source movement happened. So if you look around at open source and it's like oh IBM and later Microsoft are going to lock everything up all this cool hardware with proprietary software. We need open source software that has a bill of rights. We need a license like the GPL. This is coming for AI and AI infrastructure because similar reasons. So this is why having like seeing your projects and what you're working on locally is so important. And so okay this is really important. This is actually probably the most important segment this week. software routers. Building a software router is different in 2026 than it has been since like the late 90s. Like nerds in the late 90s and early 2000s, like building your router easy. Cisco got their start on x86. They had x86 machines that were, you know, like Pentium 2s and like some of the best Cisco hardware ever was a Pentium 2 for like the control plane plus a little bit of special hardware and that was amazing. So, fast forward to 2026. I am not bullish on pure x86based routers. And the reason for that is because we need hardware acceleration and close to the network side of the wire. I don't know exactly what that looks like. Maybe it turns into, you know, surplus like Bluefield one and two GPUs. Probably not. But having some programmability on the nick is good. There are a lot of people that share um similar thinking as me. There's a guy working on an FPGA uh based router and that's a path. U there's also like NXP semiconductor. NXP semiconductor has some SDK stuff to work with this but it's like very bad and like early 2000s and AI can kind of vibe code that to more modern Linux kernel versions. The modern switches like if you have like a Broadcom um Ethernet switch and you open it up and it's like oh there's an x86 CPU in here. The x86 CPU is just part of the control plane. The x86 CPU is telling this hugely complicated fabric chip what to do. And the fabric chip is doing the thing in hardware. So it's extremely low latency, extremely jitter, uh extremely low jitter, um extremely trouble-free. And a lot of the Linux-based routers depend on the CPU to do everything. So the CPU has to pick stuff up off of one network card, look at it, do something with it, and then deposit it on another network card or another interface. And the operation to do that is glacial at best compared to what you can get from hardware acceleration. So I think that I'm looking for options, if there are options to connect the dots on that sort of thing. Now, I have an absurdly expensive A30X Bluefield nick that's 100 Gbit um from Nvidia, and I'm delighted that I can turn that into a low jitter 100 Gbit router. That is fantastic. That is not something that is good or appropriate or affordable for the home lab. That's just cast off enterprise crap that, you know, I have to deal with for day job. So, I would love to build a router that has hardware acceleration. Some of the hardware acceleration could come with RDMA uh or find a chipset that works really well for this. The guy that's working on the NXP based router is already clued into all this and has the right mindset. But the reason why that's exciting versus just build this thing that runs at 10 Gbit is that even at 10 Gbit um the non awesomeness of that approach starts to show. you can get a 10 GB, you know, like consumer grade or even like a micro tick router that is a much better option, lower jitter, higher throughput than something that you can build, you know, around like this kind of a platform even with it relatively high clock speed in the, you know, 5 to 10 Gbit internet connection range just because the hardware is doing the switching. But there have been security vulnerabilities um in Microick and in other routers that I think would be uh useful to sidestep so that we can build our own, you know, routing platform that goes beyond what you can get in PC land. And that might take the form of a PCIe nick. like maybe it makes sense for me to commission the creation of say a four port, you know, four 10 Gbit ports on a nick with a switching fabric and some control there of what's, you know, what's going on with the switching fabric and the security and other aspects. The other nice thing is like intrusion detection and intrusion detection in the 5 to 10 gigabit range while not ruining your latency. It's also a little bit of a challenge to do. So, that is probably going to be a longterm project and in the short term it's probably going to be well you can use a really high clock speed AM5 CPU like the six core 9600X because you can you know it's got a relatively high clock speed with a reasonably good 10 GB nick. You can basically build a 10 gigabit platform that works really well. It used to be in the olden days that you could take an old crappy computer from the dawn of time and build a perfectly reasonable 1 gigabit or you know 500 megabit internet connection router and it would not significantly impact your latency make your uh like jitter under load or anything like that worse. That is not really the case anymore with fiber optic internet connections and everything else is not as good as it could be. It's decent but it's But, you know, having the four to 5 GHz CPU to pick up those packets and then drop them off on another nick uh does help negate that somewhat, but that's kind of a nice computer to use for a router, too. So, it's a little bit of a it's a little annoying, but there's never been a better time to do softwarebased routing than 2026 because of the router ban and everything else. the forbidden router as a VM on a fast platform. you know, Minis form and others have their fast NAS platforms now where it's a lot of CPU horsepower and they can also make a good forbidden router VM and the forbidden router VM with hardware acceleration can be really nice and you know you can like again you're stacking bad things when you're using a softwarebased nick but some of those nicks especially the Intelbased nicks you can do SROV single root IO virtualization so you have a paraverirtualized nick or you have um hardware that you're passing through a paraverirtualizer hardware pass through are really good in terms is not making your latency worse versus the software emulated, you know, E1000. That is not fabulous for a lot of reasons. Uh I may try to get out to visit uh Patrick Kennedy at Serve the Home because he's got the hardware that I can use to show you how bad the forbidden router can be when you have softwarebased nicks versus hardware accelerated nicks. Like it's kind of a big deal. Anyway, I've rambled enough about that for now, but I do want to refresh those videos. also in colonel 7 NTFS natively NTFS and so um one of

the things you know we did dual booting due it's like ah Windows so um with a good NTFS driver it is now possible for you to have a shared Steam library so you can dual boot between Windows and Linux but not have to duplicate your games because the new NTFS driver is much more reasonable than the old one I think that's a piece of content that I'm working on if you would like to preempt me and make a post about testing that with Colonel 7 especally especially on Cashio OS and you're in a dual boot situation. I would love to hear uh a write up of your experiences on that. So yeah, good job. All right, that's enough for Linux Weekly this week. I'm wonder level one. Thanks for hanging out. Thanks for chatting. I've got a little bit of tea left, but not a ton. And uh still pretty excited for the future. All right, I'm signing out and if you can, you know, if you got any questions or I missed anything or you want to contribute something to next week's episode, uh, we have the forum. All right, I'm signing out and I'll see you there.