llamabench.ai — crowd-sourced llama.cpp result database (per-model decode/prefill rows)

Crowd-submitted llama.cpp runs. Each row gives decode tok/s (tg128), prefill tok/s, backend+build tag (e.g. 'llama.cpp b9935 · FA on'), quant, KV-cache dtype, ctx depth, speculative-decoding flag, and reporter handle. Per-request batch-1 single-stream decode; prefill is the site's reported figure. n=1 per row unless the model page shows more. Lower confidence than a curated scoreboard but a real measurement, not a spec. Accessed for A100-PCIe-40GB (10 rows), A100-SXM4-40GB (2 rows), W7900 (1 row).