GMKtec EVO-X2 Performance
We wanted to look at this system but really we wanted to look at the new AMD Ryzen AI Max+ 395 in the context of AI performance, APU performance, and gaming performance.
AMD Ryzen AI Max+ 395 AI Performance
In the video, we showed two different models. The first was a Llama 3.3 70b Q6_K Instruct model. We ran this on both the CPU and GPU, and just the GPU partitioning the memory into 64GB for the CPU and 64GB for the GPU.
Running some of the Llama 3.3 70B Q6_K on the CPU, we were down to only around 2 tokens per second. Offloading it entirely to the GPU, a task that required over 50GB of GPU memory, got us up to 3.7-3.8 tokens/ second. Simply asking “What is ServeTheHome?” gave us a great response in either case, just the huge integrated GPU was faster. That 50GB+ figure is important since it takes you out of even the class of GPUs like the NVIDIA L40S. So running on the GPU-only with enough memory meant that we could fit the model whereas we would not have been able to with smaller GPUs.
Moving to a much smaller model Qwen3-32B we also went way faster at 10.4-10.5 tokens/ second answering the same question. With that model, and other smaller models, we got very poor answers that were unusable. One of the smaller Phi models gave something that was hillarious as a result, but I will not post here (although many press and analysts at the AMD Advancing AI 2025 event heard the story.)
A lot of folks focus solely on speed, but that was a simple prompt that gave a great answer on a larger 70b model, but absolute nonsense on many smaller models. Really, that is the power of the Ryzen AI Max+ 395 with 128GB: running large models that were previously only the domain of very high-end GPUs that cost $6000+ each.
Also, we should note that this article/ video we actually finished off a few weeks before publication. Since then, AMD said that ROCm 7 and beyond were going to support RDNA 3.5 and that the idea with the Max+ 395 is that it is the notebook (and mini PC) chip to develop on. It feels safe to assume that in the future, this is a platform that will get faster.
Geekbench AMD Ryzen AI 9 HX 370 vs Ryzen AI Max+ 395 Comparisons
We had Alex pull a few Geekbench comparisons just to show some of the CPU versus the AMD Ryzen AI 9 HX 370 like what we found in the Acemagic F3A. Here is the Geekbench 5 CPU comparison:
Here is the Geekbench 6 CPU comparison:
On the CPU side, we still get Zen 5, but we get a nice uplift over the HX 370. That would be exciting, but when we switch to the GPU, things are totally different.
Here is the Geekbench 6 GPU comparison:
That is the power of the 40 compute unit RDNA 3.5 GPU. To be clear, that AMD Radeon 890M was very good. AMD just built a bigger GPU and it is being fed here with LPDDR5X instead of DDR5.
Here is the Geekbench AI 1.3 comparison:
In terms of AI, the AMD Radeon 8060S integrated is just much faster. That is part of what makes this such a great AI system.
A Word on Gaming
We typically do not publish gaming testing, but I noticed that the system was running CS2 at around 120fps on a 4K screen. Also, it was not scaling down the visual fidelity. As such, I thought, maybe this is worth just testing a bit to see if it could be used as a higher-end gaming system as well. The easy answer is that generally gaming performance fell between the RTX 4060 desktop and the RTX 4070 notebook like we found in the ASUS ROG NUC.
Lowering the resolution a bit to 1440p, I actually thought this was very playable for competitive CS2. Sam, who is our cameraman was new to CS2 as more of a console gamer, and thought it was a great experience. I am too old to be great at CS2, but I did a few Premier games and got around 26K elo for around a 10,000th spot in the leaderboards (out of a few million.)
Unlike Sam, I can feel the difference in framerate and it matters. It might be the difference of one more kill in a game, but at higher-levels even at this top ~1% of CS2 Premier, there is still a benefit to a big dGPU over this integrated graphics. Still, I tried a few standard competitive games on it, and managed to walk away with two 1v3 clutches in one night. Those are pretty rare where a solid conversion rate is 10% or so. This was the first integrated GPU where I felt like I could still have fun with it, and not look at my hardware and think “if I had a better system I would have gotten that one.” Maybe in 2-3 years when I slow enough where I simply do not have moments where I can do it anymore, this will be a perfect system.
All that said, I do not think you buy this as a gaming system. We have another project we recorded most of a few weeks ago using the GeForce RTX 4060 and 4070 and I think if you just want to game in a smaller form factor PC, those can be cheaper options. Perhaps for the mobile market this integration for gaming can make sense. On the other hand, as cool as I found the gaming to be, the feature that blew me away in terms of value in this $2000 system was the AI capabilities.
Let us get to power next.
It is a great system, but it’s been a while since I’ve seen such an ugly enclosure on a modern computer.
Pls inform me cost of evo x2
Make it 10GbE, quieter, and not an ugly AF box and it’ll sell like sno cones in Hawaii
what makes it ugly compared to a square black box?
I’d agree with others this is a horror show on the eyes. At least a square black box is easy to hide. This doesn’t have clean lines and 2 tone with RGB
In my opinion the Ryzen processor is very nice. At the same time, I’m not interested in a weird looking mini PC that is difficult to service. On the other hand, if there were a mini ITX motherboard with this processor, I could see placing four in a suitable 2U rack-mount case.
I like it aesthetic much more than a box, give me interesting asymmetrical every time over “it’s a box.
I differ on the claim “it’s meant to be vertical” , the feet were thought after naming ports, if you want to read it the way it’s meant to be read, this is a horizontal box.
They probably decided it’d have smaller footprint on desk or against a wall while vertical aaaand, “lets add feet on edge”
Where did the get the designer for this thing from? Such a great machine in such a poor housing! Looks like a 1987 clone pc!
Ugly?!? It’s a dam computer it’s a new platform . Why is there always something to bitch about. Couldn’t complain about it performance so this is you complaint???
WOW
Great review, thanks a lot! But what about image and video gen? Is that possible with 96GB RAM dedicated for the GPU?
(a) Typo on this page (bargin, rather than bargain)
(b) You mention Windows on GB10. Is that actually confirmed or conjecture?
So much hate on the design of the case. I mean, yeah, it’s a little weird looking, but I’m fairly certain I’ve seen worse. I’m indifferent to case aesthetics, myself. Form follows function, as I see it. If the thing works as the manufacturers intended, that’s all that matters. RGB, funky stepped housing design, all that’s secondary to the actual functioning of the unit. Yeah, there’s a few places where they could have done better, but if it works and does what its designers set out to accomplish, is the visual aesthetic really THAT big an issue?
The box is meant to stay vertical. If you let it stay horizontal, you are going to clog the cpu ventilation holes. I have been working on the evo-x2 for about two weeks and I am quite happy with it.
The FA-EX9 looks a bit more promising as a workstation, no 10gbe but 128GB 8000mhz RAM like the X2 and it has oculink (seems it’s shared with the second nvme port?). It’s starting to popup on aliexpress but I haven’t found a full review yet.
Oculink seems like a great choice until you realize how much an additional gpu enclosure and gpu costs. You might end up with another 2k.
Frame.work has an mITX board with 128Gb of RAM for 2000$, shipping end of the year if ordered now. Jeff Geerling has a video where he put them in a miniRack. An miniPC with triple fans isn’t really what I want on my desk. As an enthusiast I’d rather put it in a jonsbo N5 with lots of storage as an all-in-one NAS/compute server.