#3179 Community benchmark on AMD Radeo RX 7900 XTX

Description

Measured on 2026-10-19 with `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run.

## Hardware and software

- GPU and VRAM: AMD Radeon, 24 GiB VRAM
- CPU: AMD Ryzen 5 5600 6-Core Processor (6 cores, 12 threads); installed RAM: 92.2 GiB; storage: not measured
- OS: Linux 7.0.0-38-generic; GPU driver: 6.19.14.31400100; CUDA: not read
- Strata: project 0.1.45, commit 477ac13c; engine 0.1.45
- Background workloads and power limits: not recorded

## Model and configuration

- Model and quantization: qwen3.8-flash-next-iq3_s
- Context: 655360 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking on for the suite; temperature 0

## Method

Run by `python tools/strata_bench.py --long --think` against the HTTP API of a Strata server (already running); model loading is not timed. One warm-up chat and a 1K prompt are discarded first. Every prompt starts with a unique tag, so the conversation cache cannot help the fresh-prompt tests.

- short chat: the three questions of `tools/ab_engine.py`, up to 100 new tokens, 3 runs
- 4K / 16K prompts: generated word lists sized by a calibration request, 24 new tokens, 3 fresh prompts each
- agent session: system prompt and four tools, then 5 turns that each add about 1500 words of a project file and continue from the model's own earlier answers (the way agent clients resend a conversation), 64 new tokens per turn, 3 sessions; the cache numbers come from the answer's usage (`cached_tokens`) and timings (`prompt_n`)
- concurrency (when run): N requests sent at the same moment, total tokens over the wall time of the burst, latency over all requests

## Results

| Configuration | Actual prompt tokens | Reused tokens | Generated tokens | Runs | Prompt tok/s median and range | Decode tok/s median and range | TTFT seconds median and range |
| --- | ---: | ---: | ---: | ---: | --- | --- | --- |
| short chat | 86 (83-91) | 0 | 100 (100-100) | 9 | not measured | 76.3 (66.1-82.0) | 0.62 (0.61-0.63) |
| 4K prompt, fresh | 4059 (4059-4059) | 0 (0-0) | 24 cap | 3 | 1144.9 (1144.5-1145.0) | not measured | 3.57 (3.57-3.58) |
| 16K prompt, fresh | 16098 (16098-16098) | 0 (0-0) | 24 cap | 3 | 1514.2 (1512.2-1516.3) | not measured | 10.66 (10.65-10.68) |
| 32K prompt, fresh | 32089 (32089-32089) | 0 (0-0) | 24 cap | 3 | 1519.4 (1518.7-1520.4) | not measured | 21.16 (21.15-21.18) |
| agent session, follow-up turns | 14517 (14514-14529) at the last turn | see below | up to 64 | 3 | see below | 69.7 (61.0-89.3) | 2.87 (1.88-3.11) |

Ranges are the lowest and highest of the runs; `Runs` counts requests that returned a token.

### Agent session (conversation cache)

Cache hit over the follow-up turns: 68.4 % (68.3-68.4); earlier-turn tokens read again per session: 343 (341-344) (earlier turns that the cache did not hold; the 13556 (13550-13566) new tokens of the files are read once); first turn (nothing cached) first token: 1.41 s (1.34-1.41).

| Turn (run 1) | Prompt tokens | Reused | Read | Prompt tok/s | First token (s) | Decode tok/s |
| ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| 0 | 645 | 0 | 645 | 456.8 | 1.342 | 71.5 |
| 1 | 3779 | 640 | 3139 | 1109.5 | 2.859 | 72.2 |
| 2 | 5292 | 3774 | 1518 | 813.1 | 1.892 | 61.7 |
| 3 | 8715 | 5287 | 3428 | 1117.1 | 3.101 | 63.8 |
| 4 | 11616 | 8710 | 2906 | 984.6 | 2.871 | 64.0 |
| 5 | 14529 | 11611 | 2918 | 1015.1 | 2.902 | 69.5 |

Per-request data: `results.json`. Engine log lines from the start: `engine-log-excerpt.txt` (when this run started the server).

## Correctness and limitations

Not measured: this is a speed suite. The conversation uses the model's own replies, so token counts differ a little between runs. Machine details come from the server, `nvidia-smi`/`rocm-smi` and the OS; background load and power limits are not recorded. Total time of the suite: 188 s.