| Name | Source | Status | Docs |
|---|---|---|---|
| ROCm | ROCm, rocBLAS | OK | readme |
| llama.cpp | llama.cpp | OK | readme |
| ComfyUI | ComfyUI | OK | readme |
| VLLM | VLLM, triton | OK | readme |
flowchart TD
rocm-src[docker.io/rocm/dev-ubuntu-24.04] --> rocm[docker.io/mixa3607/rocm-gfx906]
rocm --> llama[docker.io/mixa3607/llama.cpp-gfx906]
rocm --> torch[docker.io/mixa3607/pytorch-gfx906]
torch --> comfyui[docker.io/mixa3607/comfyui-gfx906]
torch --> vllm[docker.io/mixa3607/vllm-gfx906]
Changing smcPPTable/TdcLimitGfx 350 => 150 reduced the hotspot by 10+- degrees with almost no drop in performance in vllm (table in vllm)
$ upp -p /sys/class/drm/card${GPU_ID}/device/pp_table set --write smcPPTable/TdcLimitGfx=150
Changing smcPPTable.TdcLimitGfx of type H from 330 to 150 at 0x1fe
Committing changes to '/sys/class/drm/card1/device/pp_table'.
All software tested on Lenovo RD450X with 256G mem and 2x MI50 32G (x16 + x8). For cooling gpus used AMD Instinct MI50 blower fan adapter (thingiverse).
cd /opt/rocm-6.4.1/bin
apt update
apt install -y rocm-validation-suite
echo 'actions:
- name: gst-581Tflops-4K4K8K-rand-bf16
device: all
module: gst
log_interval: 3000
ramp_interval: 5000
duration: 15000
hot_calls: 1000
copy_matrix: false
target_stress: 581000
matrix_size_a: 4864
matrix_size_b: 4096
matrix_size_c: 8192
matrix_init: rand
data_type: bf16_r
lda: 8320
ldb: 8320
ldc: 4992
ldd: 4992
transa: 1
transb: 0
alpha: 1
beta: 0' > ~/gst-581Tflops-4K4K8K-rand-bf16.conf
./rvs -c ~/gst-581Tflops-4K4K8K-rand-bf16.conf