NVFP4 모델이 추가됐다고해서 bench.sh 한번 돌려봤습니다.
메모리 용량 보더니 --prefill auto:32768 추천해줘서 적용했고,
max-context는 권장사항은 64k 였으나, 128k로 설정하고 진행했습니다.
소스코드 보니 avx-512가 아이스레이크서 부터 지원이라 좀 아쉽기는 합니다. <- 스카이레이크부터 지원해줬으면 하는 바램이 ㅠ.ㅠ
주력은 swift 1.5 iq3_xxs 또는 s로 쓰다가 리뷰용으로 쓰면 괜찮겠다 생각듭니다.
Community benchmark on NVIDIA GeForce RTX 5060 Ti + NVIDIA GeForce RTX 5060 Ti
Measured on 2026-10-18 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: NVIDIA GeForce RTX 5060 Ti + NVIDIA GeForce RTX 5060 Ti, 32 GiB VRAM
- CPU: Intel(R) Xeon(R) Gold 6262V CPU @ 1.90GHz (24 cores, 48 threads); installed RAM: 377.2 GiB; storage: not measured
- OS: Linux 7.0.14-19-pve; GPU driver: 615.71.09; 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-nvfp4-nvfp4-BF16
- Context: 131072 tokens; parallel slots: 1
- Settings: the install's own configuration, unchanged; thinking off for the suite; temperature 0
strata --pack pack --native qwen-nvfp4-gptq-dense.gguf --native-dense-gguf qwen-nvfp4-gptq-dense.gguf --ple-gguf ple-fp8.gguf --embd-gguf token-embd-bf16.gguf --expert-profile expert-profi
le.bin --expert-cache auto --prefill auto:32768 --spec 4 --mtp rt --max-context 131072 --kv int8 --kv-resident 32768 --spec-min-p 0.5 --adapt-every 1 --adapt-swaps 160 --adapt-decay 0.97
Method
Run by python tools/strata_bench.py against the HTTP API of a Strata server (started by the benchmark from the install configuration); 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 |
51.6 (47.6-59.5) |
0.56 (0.53-0.64) |
| 4K prompt, fresh |
4019 (4019-4019) |
0 (0-0) |
24 cap |
3 |
455.4 (454.4-458.1) |
not measured |
8.85 (8.80-8.87) |
| 16K prompt, fresh |
16058 (16058-16058) |
0 (0-0) |
24 cap |
3 |
1275.9 (1272.7-1282.8) |
not measured |
12.62 (12.56-12.65) |
| agent session, follow-up turns |
14583 (14571-14594) at the last turn |
see below |
up to 64 |
3 |
see below |
66.2 (40.3-84.6) |
8.86 (8.35-9.04) |
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: 68.6 % (68.6-68.8); earlier-turn tokens read again per session: 72 (3-73) (earlier turns that the cache did not hold; the 13718 (13717-13719) new tokens
of the files are read once); first turn (nothing cached) first token: 6.04 s (6.03-6.06).
| Turn (run 1) |
Prompt tokens |
Reused |
Read |
Prompt tok/s |
First token (s) |
Decode tok/s |
| 0 |
609 |
0 |
609 |
101.5 |
6.026 |
31.9 |
| 1 |
3741 |
620 |
3121 |
352.4 |
8.89 |
47.6 |
| 2 |
5274 |
3805 |
1469 |
176.4 |
8.359 |
56.8 |
| 3 |
8707 |
5267 |
3440 |
383.5 |
9.002 |
65.9 |
| 4 |
11631 |
8769 |
2862 |
324.4 |
8.856 |
84.6 |
| 5 |
14571 |
11672 |
2899 |
328.3 |
8.862 |
58.6 |
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-s mi and the OS; background load and power limits are not recorded. Total time of the suite: 264 s.
