#3020 Community benchmark: RTX 5090, Qwen3.8-Flash-Next IQ3_S, 512K YaRN, parallel 2

Beschreibung

## What this is

Results only, for one PC. No engine change.

- GPU: NVIDIA GeForce RTX 5090 32 GiB, driver 617.42, power limit 402.50 W, PCIe gen 4 x16 while busy. The desktop is not on this GPU (`display_active` Disabled).
- CPU: AMD Ryzen 7 9700X. Installed RAM about 93.7 GiB. OS: Windows 11.
- Strata 0.1.45, commit 477ac13c, release engine CUDA 13.0.
- Model: Qwen3.8-Flash-Next GSQ-RCO IQ3_S. Context 524288 with YaRN factor 2 (`yarn-orig-ctx` 262144). KV int8, 32768 resident. `parallel` 2 (`--batch` 2). Expert cache auto (this start: 10038 slots / 19.05 GiB). `--vram-reserve-mib` 700, left there even though startup suggested more. Conversation cache 16384 MiB / 4 slots. `--pcie-frac` 0.35, `--spec` 4, `--spec-min-p` 0.50, `--adapt-swaps` 160. Thinking off for every measured request.

Folder: `bench/results/2026-10-18-community-rtx-5090-qwen3-8-flash-next-iq3-s/`.

## What was measured

The official suite (`tools/strata_bench.py`: chat, 4K, 16K, 32K, agent session, 2 short clients) and a long-context extension at about 64K, 128K, 256K, and a full-window prompt of about 511K tokens (the word-list estimate aimed at 523976 and landed short of that). Three runs where the process finished them. Fresh prompts, restored prompts, and two prompts at once are separate rows. Speeds are the engine's timings. A needle check covers 128K, 256K, and 480K at depths 10 and 90. All six were found.

## Limitations worth reading before quoting a number

- A cache hit is a restore in the engine log plus at least 90% of the prompt reused, after both slots were flushed. An immediate resend is not used as the cached condition.
- At 64K and 128K the single-request restore works (time to first token about 0.18 s and 0.26 s). The cached pair restored only one of the two prompts. The restored request decoded at about 33 tok/s with no draft accepts while the other request reread the prompt. That cell is not two cache hits.
- The full-window snapshot was refused on a fresh process: `skip parking (physical RAM admission; need 8243 MiB plus 2560 MiB floor, 9377 MiB available)`. The 16 GiB conversation-cache setting is not the check that fired. The resend read the prompt again.
- The same full-window length inside the long process decoded at about 26 tok/s with 0 draft accepts. A fresh process, with about 193 draft accepts, decoded that length at 180 tok/s (and 236 tok/s on the reread). Those are different rows.
- Two long prefills ran one after the other. The slower time to first token is queueing, not batched decode. There is no parallel-1 arm. YaRN is on for the short rows as well as the long ones.
- `--vram-reserve-mib` stayed 700. Startup reported 119 MiB of VRAM free and suggested a higher reserve.

`long_matrix.py` in the folder is the extension. The official suite is the repository's own tool.