#2008 Community benchmark: 2x RTX 5060 Ti 16 GB (layer split), IQ2_XS, 262K, engine 0.1.43

说明

Full report in **bench/results/2026-10-11-community-2x-rtx-5060-ti-iq2-xs/** (`README.md`, `results.json`, `engine-log-excerpt.txt`).

## Hardware

- GPU: 2x NVIDIA RTX 5060 Ti 16 GB, driver 595.91.07
- Card 0: CPU-attached slot — **PCIe 3.0 x8** (engine probe 6.7 GB/s host→device)
- Card 1: B550 chipset slot — **PCIe 3.0 x4** (engine probe 3.1 GB/s)
- CPU: AMD Ryzen 5 4500 (Zen 2), 6 cores / 12 threads, no AVX-512
- Motherboard: MSI B550-A PRO (MS-7C56), BIOS A.L1
- RAM: 64 GB DDR4 in three mixed modules, linked at 2133 MT/s
- OS: Ubuntu 24.04.5 LTS, kernel 7.0.0-38-generic
- Strata: 0.1.43 (`7695859a`), CUDA 13 source build, sm_120

## Configuration

- Flash-Next IQ2_XS, 262,144-token context, layer split 27, 16,136 resident experts (21.7 GiB), int8 KV + `--kv-resident 32768`
- `--spec 4 --spec-min-p 0.55`, adaptive tier `1/160/0.97`, `--pipeline-windows 2`, experimental speed projection on, GPU vision on

## Results

`tools/strata_bench.py --url http://127.0.0.1:8081 --long --runs 3`, run against the server already serving; medians with ranges, all nine fresh prompts `0 reused`.

| Test | Prompt tok/s | Decode tok/s | TTFT s |
| --- | --- | --- | --- |
| 4,019-token prompt | 732.6 (729.1-735.6) | — | 5.51 (5.49-5.54) |
| 16,058-token prompt | 1,139.8 (1,138.6-1,142.7) | — | 14.13 (14.09-14.14) |
| 32,049-token prompt | 1,367.2 (1,352.0-1,370.0) | — | 23.50 (23.45-23.76) |
| short chat (n=9) | — | 85.3 (58.6-104.7) | 0.36 (0.32-0.66) |
| agent turns (n=15) | — | 92.8 (65.0-114.6) | 5.37 (4.36-5.74) |

Conversation cache over the agent turns: 68.8 % hit.

## Notes

- The 5060 Ti pair is already in the index, but as IQ3_XXS / IQ3_S on Xeon and i9 hosts on engines 0.1.34-0.1.39; this adds IQ2_XS on 0.1.43. Prompt rates land below those rows (911-2,087 tok/s at 4K-32K) — the gen 3 x8/x4 link and the 6-core CPU are why; decode is in the same band.
- Not tested: concurrency (`--clients`), needle recall, vision requests, prompts beyond 32,049 tokens.
- Results-only: adds that folder plus the two index rows (`bench/results/COMMUNITY.md`, `docs/COMMUNITY_BENCHMARKS.md`); no code, defaults or docs-of-record touched.