#1927 bench: community report, RX 6900 XT (gfx1030) x1 and x2 on 0.1.42 - Unsloth UD-Q4_K_XL beside GSQ-RCO IQ3_S, stock and with the gfx103x PRs

本文

## Summary

Community benchmark folder `bench/results/2026-10-11-community-rx-6900xt-ud-q4_k_xl-iq3_s-0.1.42/` from the machine of the 0.1.41 report (2x RX 6900 XT 16 GB, gfx1030, PCIe 4.0 x8 each, Ryzen 5 5600X, 128 GB, ROCm 10.0). It adds Unsloth UD-Q4_K_XL to that machine's numbers, next to GSQ-RCO IQ3_S, on 0.1.42 built with `-DSTRATA_MMQ_KQUANTS=ON` (the flag #1897 proposes for setup).

Eight arms (format x one card / layer split x stock / the three open gfx103x PRs + switches), 4K / 32K / 128K, two runs each, zero stalls:

| | prompt 4K / 32K / 128K | decode |
|---|---|---|
| one card, IQ3_S, stock -> prs | 476 / 504 / 488 -> 959 / 1,153 / 1,129 | 45-47 -> 47-53 |
| one card, UD-Q4_K_XL, stock -> prs | 471 / 505 / 489 -> 676 / 1,117 / 1,087 | 30-32 -> 32-33 |
| split, IQ3_S, stock -> prs | 464 / 766 / 873 -> 985 / 1,792 / 2,046 | 71-72 -> 75-82 |
| split, UD-Q4_K_XL, stock -> prs | 462 / 763 / 873 -> 756 / 1,748 / 1,981 | 43-47 -> 45-51 |

What the folder says beyond the table: stock prompts are the same for both formats (the dense FP16 route is the limit, not the experts); 0.1.42's stock split equals 0.1.41's on this machine (464 / 767 / 869); UD-Q4_K_XL decodes 30-35% slower (experts 53% larger, 40% fewer cache slots); the 8K / 16K one-card arms match #1897's x16 numbers from an x8 slot (1,085-1,093 / 1,121-1,122 over three starts); and the llama.cpp RDNA2 MMQ patches that help the dense 27B do nothing for Strata's expert GEMMs (+-1% prompt, measured with and without `GGML_CUDA_MMQ_Q8K` compiled into `strata_mmq`), so the K-quant MMQ flag is the lever for this format on RDNA2.

Per-arm `config.json`, `summary.json`, `results.json` (every output text) and `BUILD.txt`, plus `benchmark.py` (unchanged from the earlier reports). Paths are rewritten (`/work/stock`, `/work/prs`, `/m`, `/bench`).

Developed with an AI coding assistant; every number above was measured on 2x RX 6900 XT (gfx1030, PCIe 4.0 x8 each) / Ryzen 5 5600X, ROCm 10.0.

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