Apple M3 Ultra (512GB), llama.cpp/MLX: Llama-3.3-70B 8-bit decode tok/s (Silicon Score measured)

THIRD-PARTY MEASURED (Silicon Score Llama 3.3 70B model page + chip rows, cited by research-big-memory.md). M3 Ultra 512GB, ~819 GB/s unified memory: 8.5 tok/s on 70B 8-bit (vs 15.5 tok/s on the same model at 4-bit - the 8-bit weight set reads ~2x the bytes per token, so the halving is the bandwidth-bound signature, not a slowdown).

Gap-fill for the Oct-14 Phase-A ingest: research-big-memory.md gave both the 4-bit (15.5) and 8-bit (8.5) rows for the M3 Ultra but only the 4-bit had a record. The pair is the cleanest demonstration on this hardware class that decode tok/s tracks bytes-per-token at fixed bandwidth.