kyuz0 AMD Strix Halo toolboxes — llama-bench results.json (pp512/tg128 across ROCm 6.4.4/7.2.3/7-nightlies + Vulkan AMDVLK/RADV)
benchmark
Fields
id
sources/kyuz0-strix-halo-toolboxes-llamabench-results-jsonPublisher
kyuz0 (Strix Halo community)
Kind
benchmark
Accessed
2026-10-17
notes
Machine-readable llama-bench sweep on AMD Ryzen AI Max+ 395 / Radeon 8060S (gfx1151): 554 runs, pp512/tg128 + long-context depths, across ROCm 6.4.4 / 7.2.3 / 7-nightlies and Vulkan AMDVLK/RADV, llama.cpp builds 9187/9193, Fedora 43. Model matrix includes llama-2-7b.Q4_0, gpt-oss-120b-mxfp4, gemma-4-26B/31B, GLM-4.7-Flash, MiniMax-M2.7, Qwen3.5/3.6. Long-context cells (pp2048 and tg32 at context depth 32768 and 65536) were filed 2026-10-17; the 2026-10-14 pass covered only the default-context pp512/tg128 cells. Re-fetched the same URL on 2026-10-17; generated_at 2026-05-18T06:42:13Z, 538 successful runs.
Notes
Machine-readable llama-bench sweep on AMD Ryzen AI Max+ 395 / Radeon 8060S (gfx1151): 554 runs, pp512/tg128 + long-context depths, across ROCm 6.4.4 / 7.2.3 / 7-nightlies and Vulkan AMDVLK/RADV, llama.cpp builds 9187/9193, Fedora 43. Model matrix includes llama-2-7b.Q4_0, gpt-oss-120b-mxfp4, gemma-4-26B/31B, GLM-4.7-Flash, MiniMax-M2.7, Qwen3.5/3.6. Long-context cells (pp2048 and tg32 at context depth 32768 and 65536) were filed 2026-10-17; the 2026-10-14 pass covered only the default-context pp512/tg128 cells. Re-fetched the same URL on 2026-10-17; generated_at 2026-05-18T06:42:13Z, 538 successful runs.
Cross-references
This record references no other records.