Speed test matrix

Qwen3.5-9B-FP8-dynamic on RTX 3090 24GB

vllm · 0.32.0 · FP8 dynamic6 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 31, 2026, 12:34 UTC

Hardware

Type: Discrete GPU
GPU: RTX 3090
VRAM: 24 GB
CPU: Intel Core i7-9700K (8C)
RAM: 32 GB
OS: Ubuntu 24.04.5

Engine

Engine: vllm
Version: 0.32.0
Quantization: FP8 dynamic
Backend: cuda
Repository: github.com/vllm-project/vllm
Build: vllm/vllm-openai:v0.32.0-x86_64

Run

Served model: bench
Revision: main
Benchmark tool: llm-inference-bench v0.8.0
Decode test: Fixed duration
Prefill mode: Measured during decode warm-up
Time per measurement: 30 s
Max output: 8,192 tokens

Launch command

# Untuned out-of-the-box baseline. Spec decode: MTP, 2 tokens lookahead. KV cache: fp8
# 1x RTX 3090 24GB (Zagitta/gpu-unlock), 420 W limit, PCIe Gen3 x16 | Intel Core i7-9700K 8C | 32 GB DDR4 | Ubuntu 24.04.5, NVIDIA driver 580.178.04
vllm serve RedHatAI/Qwen3.5-9B-FP8-dynamic --max-model-len 262144 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --language-model-only --speculative-config {"method":"mtp","num_speculative_tokens":2} --kv-cache-dtype fp8

Prefill

prompt tok/s by context length

8K context6,4181,276 ms TTFT

16K context5,9252,738 ms TTFT

32K context5,3636,021 ms TTFT

64K context4,53114,217 ms TTFT

128K context3,44137,397 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests

Context	1	2	4	8	16	32	64	128
0	96.165 ms TTFT	191.671 ms TTFT	365.9125 ms TTFT	675.1180 ms TTFT	1,185.6294 ms TTFT	1,848.4444 ms TTFT	Over capacity	Over capacity
16K	100.3270 ms TTFT	197.3481 ms TTFT	375.3805 ms TTFT	668.61,254 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
32K	95.2359 ms TTFT	194.2628 ms TTFT	360.9981 ms TTFT	612.41,531 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	94.3505 ms TTFT	185.9829 ms TTFT	341.81,336 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	85.2847 ms TTFT	172.51,288 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	75.91,629 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity

Over capacity: not run because this many requests at this context would exceed the server's KV cache.

TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.

All measurements

Latencies are medians. Total tok/s is output across all simultaneous requests.

Phase	Context	Requests at once	Status	Total tok/s	Per request tok/s	TTFT	Time between tokens	Average active requests	Completed requests	Output tokens	Measured for
Prefill	8K	1	Valid	6,418.0	—	1,276 ms	—	—	1	—	—
Prefill	16K	1	Valid	5,925.0	—	2,738 ms	—	—	1	—	—
Prefill	32K	1	Valid	5,363.0	—	6,021 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,531.0	—	14,217 ms	—	—	1	—	—
Prefill	128K	1	Valid	3,441.0	—	37,397 ms	—	—	1	—	—
Sustained decode	0	1	Valid	96.1	97.6	65 ms	10.2 ms	1.0	0	2,882	30.0 s
Sustained decode	0	2	Valid	191.6	96.9	71 ms	10.3 ms	2.0	0	5,746	30.0 s
Sustained decode	0	4	Valid	365.9	92.5	125 ms	10.8 ms	4.0	0	10,977	30.0 s
Sustained decode	0	8	Valid	675.1	86.0	180 ms	11.6 ms	8.0	0	20,246	30.0 s
Sustained decode	0	16	Valid	1,185.6	75.4	294 ms	13.3 ms	16.0	0	35,550	30.0 s
Sustained decode	0	32	Valid	1,848.4	58.5	444 ms	17.1 ms	32.0	0	55,276	29.9 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	100.3	100.8	270 ms	9.9 ms	1.0	0	3,007	30.0 s
Sustained decode	16K	2	Valid	197.3	99.8	481 ms	10.0 ms	2.0	0	5,919	30.0 s
Sustained decode	16K	4	Valid	375.3	94.7	805 ms	10.6 ms	4.0	0	11,258	30.0 s
Sustained decode	16K	8	Valid	668.6	82.3	1,254 ms	12.2 ms	8.0	0	20,054	30.0 s
Sustained decode	16K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	95.2	96.5	359 ms	10.4 ms	1.0	0	2,855	30.0 s
Sustained decode	32K	2	Valid	194.2	98.6	628 ms	10.1 ms	2.0	0	5,825	30.0 s
Sustained decode	32K	4	Valid	360.9	91.0	981 ms	11.0 ms	4.0	0	10,820	30.0 s
Sustained decode	32K	8	Valid	612.4	76.1	1,531 ms	13.1 ms	8.0	0	18,367	30.0 s
Sustained decode	32K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	94.3	95.1	505 ms	10.5 ms	1.0	0	2,826	30.0 s
Sustained decode	64K	2	Valid	185.9	94.7	829 ms	10.6 ms	2.0	0	5,576	30.0 s
Sustained decode	64K	4	Valid	341.8	86.5	1,336 ms	11.6 ms	4.0	0	10,252	30.0 s
Sustained decode	64K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	1	Valid	85.2	85.4	847 ms	11.7 ms	1.0	0	2,553	30.0 s
Sustained decode	128K	2	Valid	172.5	87.6	1,288 ms	11.4 ms	2.0	0	5,171	30.0 s
Sustained decode	128K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	1	Valid	75.9	77.0	1,629 ms	13.0 ms	1.0	0	2,275	30.0 s
Sustained decode	248K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	128	Over capacity	—	—	—	—	—	—	—	—

Over capacity: not run because this many requests at this context would exceed the server's KV cache.