Skip to content

bench/results: gfx906 IQ3_XXS (2x Instinct MI50 / Radeon VII), and the same run with --mmap-experts #1978

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
xxDoman wants to merge 1 commit into from
Closed

bench/results: gfx906 IQ3_XXS (2x Instinct MI50 / Radeon VII), and the same run with --mmap-experts #1978

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 60 additions & 0 deletions ...y-2x-instinct-mi50-32gb-qwen3-8-flash-next-uncensored-iq3-xxs-low-ram/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
# Community benchmark on AMD Instinct MI50 32GB + AMD Radeon VII 16GB

Measured on 2026-10-11 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 Instinct MI50 32GB + AMD Radeon VII 16GB, 48 GiB VRAM
- CPU: 12th Gen Intel(R) Core(TM) i5-12400F (6 cores, 12 threads); installed RAM: 61.6 GiB; storage: not measured
- OS: Linux 7.0.0-38-generic; GPU driver: 7.0.0-38-generic; CUDA: not read
- Strata: project 0.1.43, commit not a git checkout; engine 0.1.43
- Background workloads and power limits: not recorded

## Model and configuration

- Model and quantization: Qwen3.8-Flash-Next-Uncensored-IQ3_XXS
- Context: 131072 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0

## Method

Run by `python tools/strata_bench.py --clients 1` 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 | 46 (43-51) | 0 | 100 (100-100) | 9 | not measured | 41.2 (34.8-53.0) | 0.69 (0.58-0.76) |
| 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 473.6 (375.1-477.0) | not measured | 8.52 (8.46-10.75) |
| 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 543.9 (543.2-546.1) | not measured | 29.56 (29.45-29.60) |
| agent session, follow-up turns | 9038 (8988-9045) at the last turn | see below | up to 64 | 3 | see below | 44.7 (33.7-53.2) | 4.76 (4.51-4.95) |

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: 70.7 % (70.7-70.9); earlier-turn tokens read again per session: 73 (4-73) (earlier turns that the cache did not hold; the 8206 (8204-8207) new tokens of the files are read once); first turn (nothing cached) first token: 3.43 s (3.42-3.45).

| Turn (run 1) | Prompt tokens | Reused | Read | Prompt tok/s | First token (s) | Decode tok/s |
| ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| 0 | 609 | 0 | 609 | 180.2 | 3.416 | 35.7 |
| 1 | 2263 | 622 | 1641 | 366.7 | 4.511 | 33.7 |
| 2 | 3956 | 2314 | 1642 | 351.2 | 4.711 | 38.6 |
| 3 | 5662 | 4020 | 1642 | 344.4 | 4.804 | 37.0 |
| 4 | 7368 | 5655 | 1713 | 349.1 | 4.945 | 44.4 |
| 5 | 9045 | 7403 | 1642 | 340.4 | 4.862 | 52.2 |

### Concurrency: 1 clients

Total 32.5 tok/s (29.9-32.6); first token p50 0.832 s, p95 0.941 s; request latency p50 3.932 s, p95 4.239 s; errors 0. Server parallel slots: 1 (requests beyond that wait their turn).

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: 269 s.
Loading