#2718 Community benchmark: 4x RTX 3090 24 GiB (250 W each) - Unsloth UD-IQ4_XS at 262,144-token context

Description

## Summary

Community benchmark report for **4x NVIDIA GeForce RTX 3090 24 GiB** running **Unsloth Dynamic UD-IQ4_XS** (Qwen3.8-Flash-Next) at the **262,144-token native context**, measured with `tools/strata_bench.py --clients 1 --long` (3 runs per test, thinking off, medians with ranges; suite total 131 s).

| Test | Prompt tok/s (median, range) | Decode tok/s (median, range) | TTFT s (median, range) |
| --- | --- | --- | --- |
| short chat (n=9) | not measured | 113.3 (88.8-153.1) | 0.24 (0.22-0.26) |
| 4K fresh prompt (4,019 tok) | 1313.3 (1249.7-1313.8) | not measured | 3.08 (3.08-3.23) |
| 16K fresh prompt (16,058 tok) | 2393.8 (2359.3-2398.1) | not measured | 6.73 (6.72-6.83) |
| 32K fresh prompt (32,049 tok) | 3240.1 (3234.1-3248.9) | not measured | 9.92 (9.89-9.94) |
| agent session follow-ups | ~1000 mid-session | 127.1 (81.5-139.5) | 2.76 (2.05-3.19) |
| concurrency, 1 client | 75.5 total (69.9-79.4) | - | p95 0.335 |

## Hardware / software

- 4x RTX 3090 24 GiB, **each power-limited to 250 W** on the host (factory 350 W / 420 W for GPU1); GPU0 PCIe x16, GPU1-3 x4, no NVLink; layer split auto (0-11 / 12-26 / 27-35 / 36-47)
- i7-13700K, 44 GiB RAM, Debian 12 LXC on Proxmox VE, driver 595.84, CUDA 13.2, Strata 0.1.44.1 (commit 296b59d)

## Model / configuration

- Unsloth Dynamic UD-IQ4_XS (~94 GB, 3-shard GGUF), BF16 mmproj vision encoder, native-experts pack, MTP draft enabled (`--spec 4 --spec-min-p 0.5`)
- 262,144-token context, `--kv int8`, `--batch 2` (2 parallel slots), `--expert-cache auto`, `--resident-budget-gib 20` pinned RAM tier: **89% of experts resident at load**, remaining ~35.4 GiB read from a CIFS network share through the file cache (decode expert hit rate 99.7-99.8% during the suite, <=0.1% cross-GPU PCIe reads)

## Limitations

- Speed suite only; no needle/recall check was run
- The `--clients 1` arm is effectively solo: a single request occupies one slot alone and keeps MTP; with shared batch windows MTP is disabled
- 250 W per-GPU power cap lowers sustained boost clocks (~1.73-1.80 GHz observed); expect somewhat better prompt throughput on machines at stock power
- Chat decode range is wide (88.8-153.1) due to MTP draft-acceptance variance on short chats; the agent-session median (127.1) is the more stable solo-decode figure

Per-run data: `results.json`. This PR adds only `bench/results/2026-10-15-community-rtx-3090-qwen3-8-flash-next/` (README.md + results.json).