#1378 bench: community report, RX 6600 8 GB (gfx1032), EPYC 7232P, IQ3_XXS 128K on Linux

Beschreibung

## Summary

Community benchmark for an AMD Radeon RX 6600 8 GB (gfx1032) on Linux: EPYC 7232P (AVX2), 128 GB DDR4, ROCm 7.2.4, Strata v0.1.40.2. It runs Flash-Next IQ3_XXS at 128K context with KV streaming and MTP, and serves a coding agent every day.

## What changed

Adds `bench/results/2026-10-07-community-rx-6600-gfx1032/README.md` (one file, no binaries or logs):

- Cold 8.6K prompt: **263-266 tok/s** prompt and **~24 tok/s** output on 0.1.40.2 with `STRATA_HIP_PROMPT_F16=1`. That is +54 % prompt speed over 0.1.39 on the same card.
- `STRATA_PREFILL_CPU_SHARE=auto` A/B: -25 to -33 % prompt time for follow-up chunks under 1,024 tokens. This is the first AMD measurement of that switch.
- VRAM layout on an 8 GB card: 985 expert cache slots, 620 MiB free with everything loaded.
- gfx1032: the gfx1030 build works with `HSA_OVERRIDE_GFX_VERSION=10.3.0`. Only `setup.py`'s `amd_gpus()` needs a two-line mapping; the report includes it.

## Extra Notes

One machine, 2 runs per configuration, 8.6K prompts only (no 32K/128K ladder yet).

🤖 Generated with [Claude Code](https://claude.com/claude-code)