A100 40GB SXM4 Llama-3-70B Q4_K_M decode (sibling-derived estimate)

DERIVED, not measured. Basis: closest measured sibling A100-80GB SXM does 22-24 tok/s on 70B Q4_K_M (XiongjieDai RunPod llama.cpp measurement); the 40GB part has ~24% less bandwidth (1,555 vs 2,039 GB/s) -> ~17-19 tok/s expected. Recorded at midpoint 18 to make the derivation explicit; treat as a roofline-style estimate, not a run.

Bandwidth-scaled from a measured sibling, per research-used-datacenter.md. Exists so the 'what does a $2,499 SXM4 module do on 70B' question has an honest estimated row rather than a fabricated measured one.