SOTAAZ: llama.cpp KV cache quantization measured on one A100, Qwen3-8B

Third-party measurement, read 2026-10-18. Rig: one A100 80GB PCIe, driver 535, llama.cpp ggml-org master commit 69320fe (2026-09-02), CUDA build for sm80, Qwen3-8B Q4_K_M (4.68 GiB of weights), -fa 1 -ngl 99, llama-bench two-run averages. KV cache types f16, q8_0 and q4_0. Reports measured KV cache size at 32K context: f16 4.38 GiB, q8_0 2.32 GiB, q4_0 1.22 GiB. Decode at 64K depth: f16 81.6 tok/s, q8_0 44.7 tok/s (55 percent of f16), q4_0 41.2 tok/s (50 percent). Perplexity on wikitext-2, 3 chunks at 32K: f16 8.41, q8_0 8.41, q4_0 8.49 (plus 1 percent). A 4 October 2026 update says q8_0 and q4_0 hold at 56 and 53 percent of f16 at 64K on a later build. Raw JSON and the harness are linked from the post.