Halo Strix Features and What to Know

GMKtech this is the Evo X3. It's a Strix Halo machine, 128 gigs of combined LPDDR5 memory now in a completely different package. You know what was really surprising? I was at AMD's Advancing AI and the GMK tech Evo X2 was basically everywhere. This is an update. It's got a special feature. We're going to take a look and then I'm going to tell you about some things that I'm building with it because that's how we roll. If you been living under a rock, the pricing situation is sort of crazy. Although I think we've peaked and the finance bros maybe haven't caught up with that yet because it's not that things are slowing down. It's that things can't grow any faster. So to be clear about that.

But that may give some of the finance bros pause. I don't know. The reality right now is that GMKtech has one of the most affordable options if you are looking for a Strix Halo machine. It's over a thousand dollars less than other competing Strix Halo machines with the same configuration. This is of course the 128 gig configuration. I took a look at both the 64 and the 128 gig configuration for Strix Halo and a lot of our machine learning enthusiasts and college professor members of the forum, hello and welcome, are still buying a lots and lots of machines like this. This is a Strix Halo of a different design, different configuration. And look, GMKtech, they put their little sticker on the network to tell you, "Ah, this has got Windows. You want to set up

Windows without the network plugged in cuz Microsoft are kind of handsy with your privacy these days." Two USB type A ports on the front, one of which is super speed and the other one is uh USB 2.0. Got a combination headphone microphone jack. On the back we have a Realtek 2 and 1/2 gigabit LAN. I know, 2 and 1/2 gigabit in this day and age, GMK Tech didn't learn from Evo X2. Although, we do have USB 4. And that USB 4 port is pretty nice. But, here's the differentiating feature, OcuLink. There's an OcuLink interface here. Are we going to do something unholy? Are we going to combine Nvidia and AMD in the same box?

Yes. We're going to put a 560 Ti 16 gig, which is probably the best bang for the buck right now on the Nvidia side cuz you get 16 gigs of VRAM. We're also going to use the Asus Turbo R9 700 32 gig. This is also a screaming good deal. But, you can mix CUDA and ROCm. It's It's nuts. But, uh you know, RDNA and another different RDNA cuz ROCm Yeah, this is the star of the show. But, this is also Yeah, there are people out there in the forum that are rocking the dual 560 Ti. And I can't fault them cuz it's pretty good. Anyway, let's take a look at this thing so we can get up to no good with uh the build and put it all together.

Now, I have to admit, when I first reviewed GMK Tech over a year ago, I was a little suspect. But, the GMK Tech machines that I picked up have been running continuously doing various things in the office to help me with this that or testing or whatever. And so far, knock on wood, they've been pretty reliable. Now, the power delivery system here has been modified a little bit. Our Strix Halo machine can use up to 140 W for the CPU memory part of the system-on-chip configuration, whatever you want to call it. 140 W is a pretty substantial uplift of what you normally expect from a Strix Halo machine. And it is at the very top end of the TDP wattage that chip supports.

However, understand that fundamentally for AI performance like inferencing, token generation, it's not a limitation of the CPU. It's a limitation of the memory. It's the memory speed. So, the 140 W is not going to help you there. It'll help you with prompt processing speed. But, it potentially, but not your token generation speed. Uh gaming performance, it may also help there, but again, gaming performance a lot of the time depends on memory bandwidth. And the extra wattage for the CPU not going to help you for memory bandwidth. For agentic type workloads, for things that use the CPU, the 16 cores here run

circles around the 20 cores that you get in the DGX Spark for CPU type operations like compute. So, if you're more on the compute side and you want to leverage more cores, 16 cores is more cores than 20 cores in that configuration. Which is interesting. It's an interesting thing that's coming out sort of late in the Strix Halo product cycle. Very nice. GMK Tech has done a good job here with the Windows pre-installation because without the network, it didn't force me to create a cloud account. And we all know that creating a cloud account with Microsoft is a terrible idea in 2026. Now, mostly I'm going to focus on Linux and I can't wait to do some Linux testing with my GPUs. But first, we will

do some benchmarks in Windows and we will look at the Windows configuration. AMD also launched lemonade and so this video is a little bit of a chance for me to talk about lemonade and the fun stuff with lemonade. So, we're going to install lemonade on this because lemonade will let you use the uh you know, the Strix Halo facilities for doing AI stuff and that's great, but also NPU like image generation, token generation, it really makes it easy and you get pretty near the frontier maximum level of performance out of the machine. See, every couple of weeks there's little improvements, little tweaks, a model gets popular, then there's optimizations and you can do that sort

of stuff. It works great. And it goes without saying as a gaming platform, this is pretty awesome. But you shouldn't spend you know, this many thousands of dollars for the gaming side of it. I'm I'm assuming that you're using these for work, be it a creative all-in-one workload that you need or um something where you need 128 gigs of unified memory. Oh, also one thing GMK Tech is doing that I noticed is they're offering z.ai memberships. Like you can get a I think it's 3 months of membership in z.ai. That's a cloud-hosted AI model thing. If you're in the US, you can check it out. But you shouldn't send sensitive information to any cloud provider, be that z.ai or Anthropic or OpenAI. You

want to be able to run it locally. Isn't that why you're buying this machine? I mean Now in terms of the rest of the hardware bundle, I mentioned the 2 and 1/2 gig Realtek Ethernet NIC, but we also have a MediaTek RZ717 Wi-Fi 7. Uh that is a version of Wi-Fi 7 that is only 160 MHz, not 320 MHz. Some of you will have strong feelings about MediaTek in general, especially if we're going to switch to Linux. Generally on Linux these days, it's pretty much fine. You may like if you're running Arch, there's a DKMS version of the MediaTek drivers that is generally more stable. Sometimes it depends on firmware, sometimes it depends on the

specific kernel version. It's also got a Crucial P310 2 TB SSD that is PCIe Gen 4. It is uh not a performance-oriented uh M.2, but it does have pretty good characteristics. You can get pretty close to that 8 GB per second read, and as long as you don't have some sort of crazy agent bug that's just writing to the thing continuously, should have a pretty nice long lifetime. It's pretty well I like that it's Crucial. Crucial's a pretty good brand. Our memory configuration, it's Micron. It's Micron DDR5 8000. So double data rate. 8000 megatransfers.

The latency is not fabulous at 143.8 nanoseconds, 140-ish nanoseconds according to our AIDA64 test, but we do have 118 GB per second of memory bandwidth. That is about as much as a quad-channel Threadripper system. But, you know, 16 cores. So, this is its claim to fame. This is how it outperforms desktop when we're talking about, you know, AI or gaming-related stuff. This is a good cooling design. It's sucking air in from the top here and exhausting warm air out the back. And you've also got another fan on this side. So, there's kind of a lot of airflow through this tower. It sort of makes sense why they would choose a 240-W power brick.

This is a solid platform for Strix Halo. And there was a lot of people out there that are interested in a platform that gives them 128 GB of unified memory. Where we are at this point in 2026, you don't really even have to pre-allocate. You don't have to say 64 gigs is going to be the system and 64 gigs is going to be VRAM. You can just have one big pool of 128 gigs. And even on Windows, that's pretty well supported. It's sort of funny because at Computex, this was one of the design flaws or problems in general with the Windows operating system that it was not handling the unified memory well. But because Nvidia is getting ready to release RTX Spark, which includes Windows on ARM and support for Nvidia

drivers on Windows on ARM, Microsoft has corrected Windows. Guess what? That benefits Strix Halo. And the very newest builds, very newest Insider builds, the very newest builds of Windows 11, you're already getting those benefits. Now, AMD for their part in Adrenalin, if we look in Adrenalin, the Adrenalin software really quick, that's the video drivers, we can see that we still have GUI options in there to pick how much memory goes where. But even if we use standard software like LM Studio on Windows, it will correctly detect, "Hey, this is a 128-gig platform." And you can load those monster models, you know, Llama 70B. Although Llama 70B is kind of old at this point. You probably wouldn't really want to load Llama 70B.

But you can score 40 to 60 tokens per second on this platform, especially if we're talking about four and six-bit quants. There's some argument about we don't want to use four and six-bit quants cuz the loss in quality. And so, a model like that, even in LM Studio, 40 tokens a second is achievable. Even on Windows. You'll get a little bit better performance on Linux, don't get me wrong. All right, so the performance of lemonade at nearly 60 tokens a second on this particular quant of Gwen 3.6, impressive, good result.

Always ask it about perfect numbers. And this result is competent. Now, OcuLink unlocks something different. As I kind of alluded to. OcuLink is a server connector. Like, what is that? It's um it's kind of a hack. It's a kludge. It's really not meant for the uninitiated. It's not a connector that's meant for mass consumption. It's PCIe in a different form factor. Like, M.2 is PCIe or PCIe slot is PCIe or you can have PCIe other ways. The USB protocol can package PCIe kind of Well, USB-C ports can package PCIe interconnects. That's Thunderbolt. Thunderbolt 4 and 5. Thunderbolt 4 is about as fast as four lanes of PCIe Gen 3.

Thunderbolt 5 is about as fast as four lanes of PCIe Gen 4 but has overhead. This is truly four lanes of PCIe Gen 4 because it is physically PCIe Gen 4. And so, you have these OcuLink cables. Now, I happen to have an OcuLink dock here. This is really just an adapter. This is a combination USB 4 and OcuLink adapter. So, sometimes you can use a USB 4 port. But, the idea here is that this would give you four PCI Express Gen 4 lanes. That's 8 GB per second or 64 gigabit, but both ways. It's symmetrical. It's duplex. And this is the fastest way to connect a GPU to Strix Halo. What's wild here though is that when Strix Halo first launched, you know, like we did a video on the Framework desktop, that

PCIe slot doesn't have an open end or it didn't. They're They fixed that cuz good job, Framework. Uh but it also doesn't support power. So, that slot only supports delivering 25 W of power. If you use it with a GPU, you're going to have a bad time because you need to be able to support 75 W of power through a slot like that in order to have a connection. Well, this solves that problem by just giving you this. You can use all kinds of eGPU docks. There's a lot of Yeah, it's popular outside the United States because again, it's not plug-and-play. There is no USB fallback. The OCuLink is sort of an interesting thing in and of itself. And if we're going to mix and match GPU families, we're going to build something

unholy that has Nvidia and AMD DNA both, like Dr. Frankenstein fusing things together, uh we're going to switch to LM Studio for the software as well because it can handle multiple runtimes. Uh Lamda CPP only recently got multi- token prediction support as well as supporting multiple different kernels for a back end. And so, we're in a little bit uncharted territory here, especially on Windows. For me on Windows, this kind of thing is very uncharted. I've been working on this kind of thing on Linux, but it's still exciting to see that Yeah, you this basically is accessible and reachable. And don't worry, we'll we'll try RDNA. We'll try that R9 700 in just a second. All right.

Why do I have this? Why What am I working on? What is the OCuLink port got me so excited for? The USB port is free. That's the first clue. So, I've got an external connection to a GPU. I can offload things to the GPU. I have them run fast. And it works even with a 5060 Ti, which costs half as much as a 5080 with the same 16 gigs of VRAM, assuming that you can find it. Or even better, the R9 700 with 32 gigs and around $1,000, $1,200, something like that. This is going to give me the best of both worlds for certain kinds of like local AI workloads. And I'm hoping, I'm trusting that in the not too distant future, the pricing on this will stabilize because the software that we

have now is going to reach all the way back into hardware like this. And as soon as hardware is affordable again, we will have hardware on which to run these things. And I want it to run as fast as and as well as it possibly can. Now, this thing the ethernet is not that fast, but even if it's 10 gigabit ethernet, it's really not fast enough to clustering. Clustering is taking this and connecting it to another one and another one. And that USB 4 interface, that's really our best option for clustering. Now, I could use a 25 or 50 gigabit ethernet adapter in an external enclosure and we could use that for clustering. That is an option and that is a very low latency option. That is a low latency option

that is even lower latency than the USB 4 connection, but Linux kernel 7.2, Linux kernel 7.2 adds some new parameters and some new improvements to Thunderbolt 4. Apple, Apple uh you know, you've got the Mac Studio and Apple has seen a lot of development around its Thunderbolt ports lately so that you can build a cluster out of Mac Studios. The problem is that the Macs like the M4 and before, they have the unified memory and the unified memory is really fast and Apple's done a lot of things to make their memory interface fast. So, token generation can be fast. The prompt processing is kind of glacial and awful on the M4 and before and the M5 is really not where it should be. To the extent that people that are actually

spending money on building this kind of a cluster on Mac Studios have been putting things like uh Nvidia's DGX Spark in front of their Mac cluster so the DGX Spark can do the prompt processing and then the actual token generation can be off to the races with the Mac cluster. And so there are options on the Linux side that actually would make proper use of something like a 5060 Ti with its 16 gigs of fast video memory plus 128 gigs of unified memory on this platform. Or even two of these connected together with USB 4. And so, that's what I'm working on and that's what's in the forum. And so, if you want to help out with that or contribute to that or point me toward GitHub repositories, definitely join me because this thing is getting pressed

into service first for help with Thunderbolt 4 networking and performance optimization and making sure the latency is really low there. So, be on the lookout for that. That's actually going to be a whole other different video putting these things together and building a little cluster out of them. I've also got a cluster right now of the Ryzen AI Halo plus the N5 Max from Minisforum as like a storage cluster thing running Proxmox. So, look out for that video. But this is something special because it's got the OCuLink for the PCIe Gen 4, four lanes of PCIe Gen 4, and it's got the USB 4 connection. Now, the Strix Halo is a platform, you only got the three four-lane interfaces. There's not an eight-lane option with this platform as far as I

know. So, we couldn't run eight lanes of PCI Express really and get anything done. So, so what is our GMK Tech Evo X3 actually doing? Well, that's here. This is a the Ryzen AI Halo, it's another Strix Halo machine, and the N5 Max. It's a NAS. Wait, aren't these also Strix Halo? It's like it's Strix Halos it's all the way down. So, the GMK Tech is running a Qwen model that's running tools that's debugging the Thunderbolt connection between these two. Why? So, these two can work together and run larger models. It's turtles all the way down. So, the thing here is that it's not like you've got something that can run completely autonomously and run amok. Quinn 3.6 is an incredible model for what it

is, but I have to build the tools. I still have to solve the problem. It still requires brain power. If you sit down and you use a model like that, maybe it can convince you that it's doing something useful or that it can convince you that, you know, you're you're building something or whatever, but it'll really quickly sort of go off the rails. It's not a frontier model. It doesn't have a ton of intelligence. It doesn't have, you know, an omniscient the level of knowledge. But, what it can do really well is run tools. And so, I put tools together to help create and tear down the Thunderbolt network interface, and I've created tools that help it understand iPerf and do that type of benchmarks,

but also InfiniBand, like well, it's not really InfiniBand, but it's like InfiniBand tooling to do testing to figure out what the performance is. And also, the skill gives it things like looking at the CPU utilization for RDMA. And so, these are things that I would normally have to run manually or cause that manually or that fail. Like, I can automate a lot of things just by programming, but things fail in weird and unpredictable ways, and then you spend a lot of time doing programming to try to compensate for the fact that something is going to fail when you can just build the tool and you can let a model say, "Okay, well, this failed.

Let's figure it out." And then the model can surface a report to me if it can't figure it out, or you can let it run the thing. And so, this is just doing a bunch of automated testing to help me with the RDMA side of things. So then, what is the AI on these doing? Or like, these are six Halo machines, you're just using the Thunderbolt interconnect with that. That's right. But, don't forget our GM KTech also has Oculink. And so, if I want to add another device to the mix and expose it over the fabric, like if I want to do NVMe over fabric experiments, then that becomes a really nice platform to do that. And so, that's what we're we'll step into in the forum. That will be uh future technology.

All right, here we've got our BIOS tour of the Evo X3. Do note that uh the balanced configuration can leave a little bit to be desired on thermals. The performance, you may need to set your fan profile for better cooling. So, keep that in mind. You can always pop in BIOS and double-check everything. 128 gigs, our SSD is showing. Everything looks great there. You can specify up to 96 gig in the BIOS, but you shouldn't generally do that anymore. That's the old way. Here's where you set your fan speeds and temperature limits.

By default, it will not go the full fan speed until you hit 95. That's too hot for me, so I like to adjust this. I like these fan settings for a more aggressive fan profile. Otherwise, you'll complain about the thermals. They do expose SR-IOV. So, you can turn that on. That's really for that OCuLink device. It is bootable off the network if you want to do that. What does it look like for thermals? It's like, well, this thing basically stays inaudible in almost every scenario. When you're using more than 100 watts, you can definitely hear fan noise. It's more whooshing of air than high-pitched whine, but you can

definitely hear the fans. HWiNFO64, for its part, does a good job reporting system temperatures and also RPMs and how much wattage the system is using and everything else. So, if you're concerned about that, I definitely would recommend downloading HWiNFO64 and using that to keep an eye on the system both in terms of power usage and power consumption, as well as what temperatures your system is facing. There is a fine dust filter mesh behind the vents here, so it should trap dust and pet hair and things like that. But, I would have liked to have seen a little bit uh more creative of uh an implementation of a dust filter or something like that in this revised

hardware. I feel like, you know, GMKtec had learned a lot of lessons with their GMKtec Evo X2 and they put them in here and it's good to see the improvements, but also keep going. Maybe we'll see a GMKtec Evo X3 uh Gorgon Halo or Medusa down the line. 192 gigs of RAM. I don't know. We'll see. But right now, this is one of the least expensive Strix Halo machines you can get. That's pretty much it for this one. I'm Wendell at Level One. If you have any questions or I miss anything, hit me up in the forum. Check out the benchmarks and everything else in the forum. As a reminder, like there's a lot more information, a lot of write-up, a lot of nitty-gritty details

in terms of performance and everything else. Is this one of the fastest Strix Halo machines that I've tested? Yes, this configuration between the wattage and everything else is one of the fastest. DDR5 8000. There've been a couple configurations that have faster than DDR5 8000, so 8000 megatransfers. So that, you know, again, the memory speed helps you out a little bit there, but for a desktop machine, Strix Halo, 128 gigs of unified memory, there you go. I'm Wendell at Level One. I'm signing out. You can find me in the Level One forums and also other people who picked up 64 128 gig configurations of Strix Halo cuz it's it's been out a while. It's just got a new physical form

factor with some improvements. All right. I'm signing out. You can find me there.