M3 Max Llama 3.3 70B Q4_K - Medium decode (llama.cpp, siliconscore ref)
6.13 tok/s (single-request decode, per-request)
Fields
id
benchmarks/m3max-llama-3-3-70b-q4-k-medium-llama-cpp-ctx32170-tgMetric
decode_tok_s
Value
6.13
Unit
tok/s (single-request decode, per-request)
Model
Llama 3.3 70B
Methodology
Silicon Score benchmarks.csv reference run. M3 Max (GPU count not published, 64 GB): Llama 3.3 70B Q4_K - Medium via llama.cpp. context_tokens=32170. Decode tok/s single-request.
Measured by
third_party
Reproducible
yes
Thermal state
unknown
notes
siliconscore reference run; source_url=https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/. Chip variant 'M3 Max (GPU count not published, 64 GB)'.
Notes
siliconscore reference run; source_url=https://www.reddit.com/r/LocalLLaMA/comments/1he2v2n/speed_test_llama3370b_on_2xrtx3090_vs_m3max_64gb/. Chip variant 'M3 Max (GPU count not published, 64 GB)'.