lhl/strix-halo-testing llm-bench summary tables — per-model pp512/tg128 on Framework Desktop (Ryzen AI Max+ 395, 128GB LPDDR5x-8000), best-backend-per-model
benchmark
Fields
id
sources/lhl-strix-halo-llm-bench-summary-tablePublisher
lhl (Framework pre-production unit)
Kind
benchmark
Accessed
2026-10-14
notes
Summary tables 'PP Performance' and 'TG Performance' choose the best backend+flags per model. llama.cpp ~b5863, ROCm 7.0 nightly / Vulkan AMDVLK+RADV, -fa 1 --mmap 0 -ngl 99. pp512=prefill tok/s, tg128=decode tok/s, per-request batch-1. Max MiB resident column included.
Notes
Summary tables 'PP Performance' and 'TG Performance' choose the best backend+flags per model. llama.cpp ~b5863, ROCm 7.0 nightly / Vulkan AMDVLK+RADV, -fa 1 --mmap 0 -ngl 99. pp512=prefill tok/s, tg128=decode tok/s, per-request batch-1. Max MiB resident column included.
Cross-references
This record references no other records.