gfx1151 support (APU Radeon 8060S; in the Ryzen AI MAX+ 395 APU with possible 128 GB RAM)

[sinni800]

Would it be possible to have this? These new Ryzen MAX devices are coming out now and you can assign up to 96 GB of VRAM to their GPU. The RAM is LPDDR5-8000 too, so quite speedy for what it is.

[zac-market]

Bump. Does anyone know if these work?

If they work I'll buy one as an AI workstation, otherwise it's no dice.

[sinni800]

Well, llama.cpp does work, so might have to step towards that

[Danik-droid]

ggml-org#10879

"AMD Ryzen Al Max+ 395 | 1271.46 ± 3.16 | 46.75 ± 0.48 | "

Vulkan definitely works, as for rocm, from what I know, only version 6.4+ somehow supports this processor - and there is no version for Windows yet. Supposedly, it already works in rocm under Linux.

Could you share how it works for you? I'm also considering this solution.

[sinni800]

Well, the Windows system requirements for ROCM here https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html say the 395+ is supported already, that's why I was asking

[YuliaS11]

I was able to build the dll on Windows by simply pointing to 6.2 directory and changing architecture to "gfx1151"

[YuliaS11]

The rough steps on Windows:

1. Make sure you have clang in your path(which should allow you to find both clang.exe and clang++.exe)
2. Make sure you have cmake in your path
3. Preferably, you have VS installed and launch from the dev console
4. Download the 6.2.4 HIP SDK from AMD, install and replace the path below with what you have
5. Follow the steps and clone the project, make a folder called build and open the dev console there

The command to run
cmake .. -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DLLAMA_HIPBLAS=ON -DHIP_PLATFORM=amd -DCMAKE_C_COMPILER="C:/Program Files/AMD/ROCm/6.2/bin/clang.exe" -DCMAKE_CXX_COMPILER="C:/Program Files/AMD/ROCm/6.2/bin/clang++.exe" -DAMDGPU_TARGETS="gfx1151"

cmake --build . -j 16