inference/research
verifiedlow · 0.50updated 2026-10-14

lhl/strix-halo-testing llm-bench — context sweeps + DGX Spark vs Strix Halo comparison (ROCm rocWMMA / Vulkan AMDVLK-RADV)

repo

Fields

id
sources/lhl-strix-halo-testing-llm-bench
URL
https://github.com/lhl/strix-halo-testing
Publisher
Lee Henderson (lhl) / Framework pre-production cluster
Kind
repo
Accessed
2026-10-14
notes
gpt-oss-120b MXFP4 tg32/pp2048 at context depths 0..32768 on Ryzen AI Max+ 395 (Radeon 8060S gfx1151), llama.cpp build 6792, TheRock/ROCm nightly 7.10.0a20251017 + RADV 25.2.4/AMDVLK 2025.Q2.1 vs DGX Spark CUDA build 6767. DGX figures transcribed from ggerganov's #16578 sweep.

Notes

gpt-oss-120b MXFP4 tg32/pp2048 at context depths 0..32768 on Ryzen AI Max+ 395 (Radeon 8060S gfx1151), llama.cpp build 6792, TheRock/ROCm nightly 7.10.0a20251017 + RADV 25.2.4/AMDVLK 2025.Q2.1 vs DGX Spark CUDA build 6767. DGX figures transcribed from ggerganov's #16578 sweep.

Cross-references

This record references no other records.