[jubilee33]

This is an interesting review. I have a Chinese strix halo box that's isnt available in the west (favm faex1) I've been able to do some ok graphical gen, or some decent agentic tasks as a fallback for when some of the APIs are overloaded during business hours, but nothing amazing for sure, and also not both at the same time. But here's the thing...it cost me 1800usd two months ago....and it's runs x86. I am struggling to see why people pay +2x more for the Arm Nvidia version, despite the slightly higher bandwidth it still does basically the same AI tasks and alot fewer high end general computing tasks... I like my box but I wouldn't find it useful enough to pay more than I did for it or get more of them and cluster for instance. Can anyone explain the allure of the Nvidia box, other than brand name?

[nightski]

It's simple, the 395+ Max Strix Halo you bought for $1800 is now a ~$4000 build (at least the AMD AI dev unit). If only we had time travel right? Either way, the Nvidia unit comes with Connect-X 7. That may or may not matter to you, but the hardware for that isn't cheap. In general the Nvidia cards also have better support for models. I know AMD is trying to catch up but anything except their datacenter cards do not seem to be getting a lot of attention.

[kingstnap]

AMD is barely trying. The 395+ Max Strix Halo was launched January 2025.

gfx1151 was not listed in the ROCm compatibility matrix for ROCm 7.2.4 [0]. This is the previous version of ROCm.

It's only finally received support in ROCm 7.14.0 [1]! It literally just started receiving support last week.

[0] https://rocm.docs.amd.com/en/docs-7.2.4/compatibility/compat...

[1] https://rocm.docs.amd.com/en/docs-7.14.0/compatibility/compa...

[mycall]

Vulkan has worked for a long time on gfx1150/1151 and works well.

[ciupicri]

Is it to me or the iGPU of Ryzen 9950X is not supported by ROCm?

[kcb]

The igpu is a tiny RDNA2 GPU. Pretty certain the 9950x cpu will outperform it in compute tasks.

[ciupicri]

What if the software was written for ROCm?

[embedding-shape]

> isnt available in the west [...] Can anyone explain the allure of the Nvidia box

The first part might answer the second one. Otherwise, the lack of CUDA and the nvidia ecosystem of tooling could also explain why it doesn't seem so interesting for AI tasks.

[icedchai]

Yep, stuff is more likely to "just work" on NVidia. Example: pytorch

In most benchmarks, the Spark is also faster at the prompt processing / prefill phase.

[jubilee33]

But strix halo boxes themselves are available, just not that one. And despite my concerns about what's said about Cuda and ROCm I have never had a problem running any model, for image or text or voice, the community has done great work in making things work. So the point of the question stands.

It's also interesting that the most high end Chinese equipment, both prosumer things like these boxes but also the Huawei professional stack is just not available in the places it would be most appreciated. Not sure if thats china tit for tat, or western "we don't want your commie hardware anyways"

But for a lot of people it sucks cause nobody should be paying 4.2k for this product. The value isn't there.

[embedding-shape]

> I have never had a problem running any model, for image or text or voice, the community has done great work in making things work

There is a whole world of other tooling and stuff that isn't just for hobbyists to run inference with ML models, but also how to do profiling, debugging and gathering data when you run distributed workloads, and so on. The nsight toolkit seems miles ahead of the competition on other platforms, as just one example.

[mycall]

Also, you can use handheld PC 395+ 128GB with battery for (semi-){0,1}offline use cases, works great for C2 of cognitive SDR radio (among plenty of other use cases).