llama.cpp Vulkan/MoltenVK on Apple M3 Ultra: Llama-2-7B Q4_0 noFA tg128 decode tok/s
115.54 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-vulkan-m3-ultra-llama27b-q40-nofa-tg128Metric
decode_tok_s
Value
115.54
Unit
tok/s (tg128, per-request, batch 1)
Model
llama-2-7b.Q4_0.gguf (TheBloke, 3.56 GiB, 6.74B params)
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #10879). llama-bench -ngl 100 -fa 0, Vulkan backend via MoltenVK on macOS. Build 2d451c8. Per-request batch-1.
Measured by
third_party
Reproducible
yes
notes
Vulkan path hits 115.5 tg128 vs Metal 92.1 (same chip, #4167) — Vulkan/MoltenVK outperforms Metal here on Q4_0 decode.
Notes
Vulkan path hits 115.5 tg128 vs Metal 92.1 (same chip, #4167) — Vulkan/MoltenVK outperforms Metal here on Q4_0 decode.