#2030 bench: RX 9060 XT 16 GiB, Flash-Next IQ3_S at 262K, engine 0.1.43 (Linux, HIP gfx1200)

本文

## Summary
Community benchmark, suite 1 with `--long`, run on 2026-10-11. As far as I can find this is the first IQ3_S entry for the RX 9060 XT: the only other 9060 XT report in the tree (`2026-10-10-community-rx-9060-xt`) is Windows with IQ2_XS/IQ3_XXS.

Hardware and config as read by the tool: RX 9060 XT 16 GiB (gfx1200, HIP engine compiled locally), Ryzen 9 7945HX, 94.0 GiB RAM, CachyOS (Linux 7.2.9-1-cachyos), engine 0.1.43, 1 slot, 262144 context, `--kv int8 --kv-resident 32768 --vram-reserve-mib 3072`, expert cache 3447 slots / 6.6 GiB in VRAM with the remaining experts in host RAM.

Medians (ranges over 3 runs): chat decode 46.0 tok/s (39.1-56.0) with first token 0.57 s; 4K prompt 779.6; 16K 1011.5; 32K 1255.6 (first token 25.6 s); agent session decode 46.0 with 68.7 % cache hits.

Three notes, so the numbers are read correctly:

- **Idle machine.** This pass was run with the desktop idle and no other inference traffic. An earlier pass on the same build was contaminated by the submitter's own agent traffic: the 32K range went from 1049-1227 to 1253-1256, and chat decode from 41.5 to 46.0. The folder above is the idle pass.
- **Update delta on this machine.** 0.1.41 → 0.1.43, both passes run the same way (with the submitter's traffic on, so comparable to each other): chat decode 34.9 → 41.5 (+19 %), 4K prompt 641 → 760 (+18.5 %), 32K 1099 → 1227 (+11.7 %). The idle numbers in this report are faster still.
- **Cross-card sanity check.** Against #1998 (RX 9070 XT, IQ3_S, 0.1.43): we are ~1.4x slower on decode (46.0 vs 65-68) and ~2.2x on the 32K prompt (1256 vs 2726), while the agent cache hit is the same (68.7 % vs 68.8 %). That looks like the card and the link, not the configuration.

Happy to re-run with different settings if a specific comparison would be useful.