Apple M2 Max 38-core (64GB), llama.cpp/LM Studio: Llama-3.3-70B 4-bit decode tok/s (Silicon Score measured)
8.8 decode tok/s (single-stream, 4-bit, 4K ctx, LM Studio)
Fields
id
benchmarks/comm-apple-m2-max-llama33-70b-4bit-decodeMetric
decode_tok_s
Value
8.8
Unit
decode tok/s (single-stream, 4-bit, 4K ctx, LM Studio)
Model
Llama-3.3-70B (dense) 4-bit
Methodology
THIRD-PARTY MEASURED (Silicon Score Llama 3.3 70B model page + chip rows, export CSV). M2 Max 38-core GPU / 64GB unified configuration, ~400 GB/s package bandwidth shared with the CPU: 8.8 tok/s on 70B 4-bit at 4K ctx. Single-stream decode; batch-1, bandwidth-bound regime.
Measured by
third_party
Reproducible
yes
notes
CREATED 2026-10-15 Phase-A residual gap-fill - sibling rows live on their own accelerator records (M1 Ultra 12.6, M4 Max 11.8, M3 Ultra 15.5); this one was stranded in comm-apple-m1-ultra-llama33-70b-4bit-decode's notes because no apple-m2-max record existed. 64 GB fits a ~40 GB 70B weight set only at low context; the 96 GB SKU is the comfortable fit. ~400 GB/s is half an M2 Ultra's 800 GB/s and decode is bandwidth-bound, so 8.8 vs the Ultra's ~9-11 est reflects KV-read narrowing, not a paradox - at short ctx weight reads dominate and the ratio is closer to the bandwidth ratio.
Notes
CREATED 2026-10-15 Phase-A residual gap-fill - sibling rows live on their own accelerator records (M1 Ultra 12.6, M4 Max 11.8, M3 Ultra 15.5); this one was stranded in comm-apple-m1-ultra-llama33-70b-4bit-decode's notes because no apple-m2-max record existed. 64 GB fits a ~40 GB 70B weight set only at low context; the 96 GB SKU is the comfortable fit. ~400 GB/s is half an M2 Ultra's 800 GB/s and decode is bandwidth-bound, so 8.8 vs the Ultra's ~9-11 est reflects KV-read narrowing, not a paradox - at short ctx weight reads dominate and the ratio is closer to the bandwidth ratio.