#2695 bench: community report - GTX 1080 Ti (sm_61), STRATA_DENSE_F32 A/B: gemm_f32_parity bitwise on cuBLAS 12.8.4.1

说明

A results-only PR: an A/B of the opt-in `STRATA_DENSE_F32=1` (#2178) on a second sm_61 card, on a cuBLAS the route had not been checked against, following [COMMUNITY_BENCHMARKS.md](docs/COMMUNITY_BENCHMARKS.md).

## What it adds

- `bench/results/2026-10-16-community-gtx-1080-ti-dense-f32-ab/`: the `gemm_f32_parity --bench` output, four per-arm transcripts, the A/B runner script, and a README in the report format.
- One list entry in `docs/COMMUNITY_BENCHMARKS.md`, one sentence in `docs/OLDER_GPUS.md` (the parity check now also passed on cuBLAS 12.9 and 12.8.4.1).

## What it leaves alone

Everything else: no engine, setup or tool change, and no default is flipped — `STRATA_DENSE_F32` stays opt-in exactly as #2178 shipped it.

## Hardware and headline numbers

GTX 1080 Ti 11 GB (sm_61, GPU on NUMA node 1, bound per #2684), 2x Xeon E5-2696 v4, driver 580.178.04, CUDA 12.8 toolkit (cuBLAS 12.8.4.1), engine 0.1.44.1 source build, Flash-Next IQ3_S, 262,144-token context, prompt chunk auto = 4,608. Same binary in both arms, arms alternated A B A B, one instance at a time, each arm started at 57-58 C.

| | `STRATA_DENSE_F32=0` | `=1` | delta |
| --- | ---: | ---: | ---: |
| ~565-token prompt, tok/s | 240.4 (232-244) | 257.0 (249-260) | +6.9% |
| ~7,047-token prompt, tok/s | 341.4 (331-356) | 375.2 (359-404) | +9.9% |
| decode, tok/s | 27-34 | 28-34 | unchanged (noise) |

`gemm_f32_parity` on this cuBLAS: **5,201 sizes bitwise equal** — same count as the P40's cuBLAS 12.4.5.8 and the P100/cuBLAS 12.9 run in #2178. The 7,040/7,041 cuBLAS kernel cliff is present on this card too (qsa query 121.7 -> 60.7 ms).

## Limitations

Speed only: no answer-identity or quality suite was run (bitwise parity is the correctness argument here). Short and 7K prompts only; one card; one cuBLAS; Linux. The gain is smaller than the P40's because this card's chunk is 4,608 tokens and cuBLAS 12.8's FP16-operand kernels sit on the fast side at several of those sizes.

Related: #2178 (the route), #2642 (the same machine's suite report), #2684 (the NUMA binding both arms ran under).