Apple M2 Max (MacBook Pro 14/16-inch & Mac Studio, 2023)

Apple-published ~400 GB/s for the M2 Max with 38-core GPU (the 30-core GPU configuration is listed at 300 GB/s). Unified-memory bandwidth, package-level, shared with the CPU. The 38-core / 96 GB pairing is the one Silicon Score benchmarks on a 70B model.

CREATED 2026-10-15 Phase-A residual gap-fill: the big-memory research file cites an M2 Max 38-core 64GB row at 8.8 tok/s on Llama-3.3-70B 4-bit (Silicon Score), but no apple-m2-max record existed to hang that benchmark on - it lived only in comm-apple-m1-ultra-llama33-70b-4bit-decode's notes. M2 Max sits one die below M2 Ultra: 12-core CPU, 30- or 38-core GPU, 400 GB/s on the 38-core part (300 GB/s on the 30-core), up to 96 GB unified. CAPACITY-vs-BANDWIDTH caveat that matters for this part: the 64 GB configuration (the one benchmarked) can only hold a 70B 4-bit model at LOW context - the model is ~40 GB of weights and KV headroom shrinks fast past 4-8k ctx. The 96 GB SKU is the comfortable 70B fit. 400 GB/s is roughly HALF an M2 Ultra's 800 GB/s, and batch-1 decode is bandwidth-bound, so M2 Max lands near half the Ultra's tok/s on the same model (measured 8.8 vs ~9-11 est on the Ultra - the gap narrows because KV reads, not just weight reads, dominate at long ctx). UNIFIED MEMORY IS NOT VRAM - same caveats as the rest of the Apple block: the advertised capacity is shared with the OS/CPU and the bandwidth ceiling serves the whole system, not a private GPU bus. tdp_w is null: Apple publishes no per-SoC power figure for the laptop or the Studio it sits in. Sources: benchmark row from Silicon Score Llama-3.3-70B page (8.8 tok/s, 38-core, 64GB, 4-bit); bandwidth/config bounds from Apple tech specs as republished in the same Silicon Score machine rows and the llama.cpp #4167 M-series table.