llama.cpp Metal on M5 Ultra (80 GPU): Llama-2-7B q8-0 pp512 tok/s (scoreboard #4167, c1d0e7a)
4910.55 tok/s (pp512, prefill batch-512, per-request)
Fields
id
benchmarks/llamacpp-metal-m5-ultra-llama27b-q8-0-pp512-c1d0e7aMetric
prefill_tok_s
Value
4,910.55
Unit
tok/s (pp512, prefill batch-512, per-request)
Model
Llama-2-7B gguf-q8-0.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. Build commit c1d0e7a (2026 Aug 25 protocol). pp512 = prefill batch 512; tg128 = batch-1 decode. Per-request, NOT aggregate. Chip: M5 Ultra 80-core GPU, 1228 GB/s unified BW, Mac Studio 2026.
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-ultra.
Notes
Maintainer-curated Apple Silicon llama-bench table (discussion #4167). First record on local-apple-m5-ultra.