llama.cpp Metal on M5 Max (40 GPU): Llama-2-7B q4-0 tg128 tok/s (scoreboard #4167, c1d0e7a)
119.92 tok/s (tg128, batch-1 decode, per-request)
Fields
id
benchmarks/llamacpp-metal-m5-max-llama27b-q4-0-tg128-c1d0e7aMetric
decode_tok_s
Value
119.92
Unit
tok/s (tg128, batch-1 decode, per-request)
Model
Llama-2-7B gguf-q4-0.gguf
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #4167, ggerganov-maintained). llama-bench -hf ggml-org/Llama-2-7B-GGUF -p 512 -n 128 -ngl 99 --delay 30, Metal backend (MTL,BLAS), full GPU offload, 6 threads. Build commit c1d0e7a (2026 Aug 25 protocol). pp512 = prefill at batch 512; tg128 = batch-1 decode. mean +/- std over default repeats. Per-request, NOT aggregate. Chip: M5 Max 40-core GPU, 614 GB/s unified BW, 128 GB.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
Maintainer-curated Apple Silicon llama-bench table (discussion #4167). First record on local-apple-m5-max.
Notes
Maintainer-curated Apple Silicon llama-bench table (discussion #4167). First record on local-apple-m5-max.