Speed test matrix

Qwen3.5-4B-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, 17:51 UTC

Hardware

Type: Discrete GPU
GPU: RTX 3090
VRAM: 24 GB
CPU: AMD Ryzen 9 7900X (12C)
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
# 1x RTX 3090 24GB (Zagitta/gpu-unlock), 350 W limit (VBIOS max), PCIe Gen4 x16 | AMD Ryzen 9 7900X 12C | 32 GB DDR5 | Ubuntu 24.04.5, NVIDIA driver 595.99.02
vllm serve RedHatAI/Qwen3.5-4B-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}

Prefill

prompt tok/s by context length

8K context10.3K795 ms TTFT

16K context9,5451,699 ms TTFT

32K context8,4403,826 ms TTFT

64K context6,8019,473 ms TTFT

128K context4,85726,495 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	146.425 ms TTFT	292.450 ms TTFT	541.859 ms TTFT	936.0106 ms TTFT	1,464.9170 ms TTFT	2,099.1224 ms TTFT	Over capacity	Over capacity
16K	110.2135 ms TTFT	209.2208 ms TTFT	388.0387 ms TTFT	533.9647 ms TTFT	666.41,161 ms TTFT	Over capacity	Over capacity	Over capacity
32K	79.5157 ms TTFT	156.9235 ms TTFT	289.0470 ms TTFT	363.0796 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	51.5255 ms TTFT	105.0364 ms TTFT	197.5686 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	31.0457 ms TTFT	61.8625 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	16.2690 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	10.3K	—	795 ms	—	—	1	—	—
Prefill	16K	1	Valid	9,545.0	—	1,699 ms	—	—	1	—	—
Prefill	32K	1	Valid	8,440.0	—	3,826 ms	—	—	1	—	—
Prefill	64K	1	Valid	6,801.0	—	9,473 ms	—	—	1	—	—
Prefill	128K	1	Valid	4,857.0	—	26,495 ms	—	—	1	—	—
Sustained decode	0	1	Valid	146.4	150.0	25 ms	6.7 ms	1.0	0	4,390	30.0 s
Sustained decode	0	2	Valid	292.4	149.5	50 ms	6.7 ms	2.0	0	8,770	30.0 s
Sustained decode	0	4	Valid	541.8	138.7	59 ms	7.2 ms	4.0	0	16,249	30.0 s
Sustained decode	0	8	Valid	936.0	121.0	106 ms	8.3 ms	8.0	0	28,065	30.0 s
Sustained decode	0	16	Valid	1,464.9	94.8	170 ms	10.6 ms	16.0	0	43,929	30.0 s
Sustained decode	0	32	Valid	2,099.1	68.2	224 ms	14.7 ms	32.0	0	62,939	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	110.2	110.9	135 ms	9.0 ms	1.0	0	3,305	30.0 s
Sustained decode	16K	2	Valid	209.2	106.1	208 ms	9.4 ms	2.0	0	6,273	30.0 s
Sustained decode	16K	4	Valid	388.0	98.3	387 ms	10.2 ms	4.0	0	11,632	30.0 s
Sustained decode	16K	8	Valid	533.9	67.3	647 ms	14.9 ms	8.0	0	16,013	30.0 s
Sustained decode	16K	16	Valid	666.4	41.8	1,161 ms	23.9 ms	16.0	0	19,981	30.0 s
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	79.5	81.1	157 ms	12.3 ms	1.0	0	2,383	30.0 s
Sustained decode	32K	2	Valid	156.9	79.4	235 ms	12.6 ms	2.0	0	4,704	30.0 s
Sustained decode	32K	4	Valid	289.0	73.4	470 ms	13.6 ms	4.0	0	8,665	30.0 s
Sustained decode	32K	8	Valid	363.0	46.3	796 ms	21.6 ms	8.0	0	10,875	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	51.5	52.5	255 ms	19.1 ms	1.0	0	1,543	30.0 s
Sustained decode	64K	2	Valid	105.0	53.2	364 ms	18.8 ms	2.0	0	3,149	30.0 s
Sustained decode	64K	4	Valid	197.5	49.3	686 ms	20.3 ms	4.0	0	5,923	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	31.0	31.6	457 ms	31.7 ms	1.0	0	929	30.0 s
Sustained decode	128K	2	Valid	61.8	31.6	625 ms	31.7 ms	2.0	0	1,851	29.9 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	16.2	16.6	690 ms	60.1 ms	1.0	0	483	29.9 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.