#2405 bench: community report, RX 9070 XT 16 GB on Windows, IQ3_XXS, engine 0.1.44.1

Beschreibung

Results-only submission: a community benchmark of Strata 0.1.44.1 on an AMD Radeon RX 9070 XT 16 GB (gfx1201) under Windows 11, measured with the built-in benchmark (suite 1, 3 runs per test). No engine code, default or documentation is changed.

## Hardware and software

- GPU: AMD Radeon RX 9070 XT 16 GB (gfx1201), Windows driver 32.0.31036.15, stock settings (power limit and clocks not recorded)
- CPU: AMD Ryzen 7 9800X3D, 8 cores / 16 threads (AVX-512 present)
- RAM: 46.8 GiB; Windows commit limit about 124 GiB (a 77 GiB page file)
- Storage: the model shards and the prepared pack on an SK hynix PVC10 HFS001TEM9X173N NVMe SSD
- PCIe: Gen5 x16; the engine's own start-up probe measured 54.9 GB/s host->device
- OS: Windows 11 (10.0.26200)
- Engine: 0.1.44.1, the prebuilt Windows HIP zip (ROCm 10.2.0a20260930)

## Model and configuration

Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS, 262144-token context, the install's own settings:

`--kv k8v4 --kv-resident 32768 --resident-experts --vram-reserve-mib 256 --spec 4 --spec-min-p 0.5 --mtp ... --max-context 262144`

Runtime state of the run: 5293 experts in the GPU cache (8.58 GiB of VRAM), 32.32 GiB of experts resident in RAM (pageable).

## Results (medians over 3 runs, the engine's own timings)

| Test | Prompt tok/s | Decode tok/s | First token |
| --- | ---: | ---: | ---: |
| short chat, 100 tokens out | - | 65.7 (60.5-79.3) | 0.39 s |
| 4K fresh prompt | 800.5 (799.4-808.5) | - | 5.04 s |
| 16K fresh prompt | 921.1 (920.6-926.5) | - | 17.46 s |
| agent session, follow-up turns | - | 68.8 (40.0-87.3) | 4.67 s |

Conversation-cache hit over the agent session's follow-up turns: 68.7%.

## Second folder: a `--kv-resident` A/B on the same machine

`2026-10-14-community-rx-9070-xt-qwen3-8-flash-next-iq3-xxs-2/` is the same machine, day and configuration with `--kv-resident 32768` removed. Prompt throughput 1.9-2.5% lower and first-token time 2.0-2.5% higher (both with non-overlapping run-to-run ranges); the decode medians are lower too (chat -4.4%, agent -6.8%) but with overlapping ranges, so those are directional only. With streaming the engine also skipped 22% fewer expert contributions (`route tail skip`), so it is the better of the two on both counts. The two folders are submitted together because they differ in exactly one flag.

## Limitations

- Speed only: no correctness, recall or needle check was run.
- One machine, one engine session, the desktop in normal use; no clock or power capture, and nothing was isolated or pinned.
- The reported commit is a local branch: upstream `main` at `296b59d2` plus two local patches (a Simplified Chinese UI translation and the Windows AMD telemetry of PRs #1434/#1152). Neither touches the engine or the model path, so the numbers are the engine's own.