Speed test matrix

gpt-oss-20b on RTX 3090 24GB

vllm · 0.32.0 · MXFP45 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 06:40 UTCEdited Oct 29, 2026, 12:41 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: MXFP4
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: none. KV cache: int4_per_token_head
# 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 openai/gpt-oss-20b --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --kv-cache-dtype int4_per_token_head --attention-backend TRITON_ATTN

Prefill

prompt tok/s by context length

8K context7,6631,064 ms TTFT

16K context5,6292,860 ms TTFT

32K context3,7338,570 ms TTFT

64K context2,22328,691 ms TTFT

120K context1,29792,068 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	146.351 ms TTFT	221.851 ms TTFT	366.382 ms TTFT	531.889 ms TTFT	898.8102 ms TTFT	1,748.1151 ms TTFT	2,849.9245 ms TTFT	Over capacity
16K	123.668 ms TTFT	199.891 ms TTFT	306.0135 ms TTFT	461.4256 ms TTFT	617.9481 ms TTFT	911.9918 ms TTFT	Over capacity	Over capacity
32K	108.2100 ms TTFT	172.1161 ms TTFT	177.3232 ms TTFT	275.4478 ms TTFT	415.0903 ms TTFT	Over capacity	Over capacity	Over capacity
64K	86.8154 ms TTFT	136.9282 ms TTFT	179.3409 ms TTFT	208.3792 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
120K	64.2310 ms TTFT	89.3506 ms TTFT	110.0660 ms TTFT	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	7,663.0	—	1,064 ms	—	—	1	—	—
Prefill	16K	1	Valid	5,629.0	—	2,860 ms	—	—	1	—	—
Prefill	32K	1	Valid	3,733.0	—	8,570 ms	—	—	1	—	—
Prefill	64K	1	Valid	2,223.0	—	28,691 ms	—	—	1	—	—
Prefill	120K	1	Valid	1,297.0	—	92,068 ms	—	—	1	—	—
Sustained decode	0	1	Valid	146.3	147.4	51 ms	6.8 ms	1.0	0	1,872	12.8 s
Sustained decode	0	2	Valid	221.8	115.6	51 ms	8.7 ms	2.0	0	6,655	30.0 s
Sustained decode	0	4	Valid	366.3	93.4	82 ms	10.7 ms	4.0	0	10,989	30.0 s
Sustained decode	0	8	Valid	531.8	77.7	89 ms	12.9 ms	8.0	0	15,955	30.0 s
Sustained decode	0	16	Valid	898.8	64.3	102 ms	15.5 ms	16.0	0	26,961	30.0 s
Sustained decode	0	32	Valid	1,748.1	56.7	151 ms	17.6 ms	32.0	0	52,431	30.0 s
Sustained decode	0	64	Valid	2,849.9	49.7	245 ms	20.1 ms	64.0	0	85,491	30.0 s
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	123.6	123.9	68 ms	8.1 ms	1.0	0	3,706	30.0 s
Sustained decode	16K	2	Valid	199.8	99.2	91 ms	10.1 ms	2.0	0	5,984	29.9 s
Sustained decode	16K	4	Valid	306.0	75.0	135 ms	13.3 ms	4.0	0	9,177	30.0 s
Sustained decode	16K	8	Valid	461.4	57.7	256 ms	17.3 ms	8.0	0	13,842	30.0 s
Sustained decode	16K	16	Valid	617.9	42.3	481 ms	23.7 ms	16.0	0	18,538	30.0 s
Sustained decode	16K	32	Valid	911.9	31.6	918 ms	31.6 ms	32.0	0	27,332	30.0 s
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	108.2	108.5	100 ms	9.2 ms	1.0	0	3,247	30.0 s
Sustained decode	32K	2	Valid	172.1	85.8	161 ms	11.7 ms	2.0	0	5,154	29.9 s
Sustained decode	32K	4	Valid	177.3	62.6	232 ms	16.0 ms	4.0	0	5,317	30.0 s
Sustained decode	32K	8	Valid	275.4	45.4	478 ms	22.0 ms	8.0	0	8,260	30.0 s
Sustained decode	32K	16	Valid	415.0	31.2	903 ms	32.0 ms	16.0	0	12,444	30.0 s
Sustained decode	32K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	86.8	87.0	154 ms	11.5 ms	1.0	0	2,603	30.0 s
Sustained decode	64K	2	Valid	136.9	67.9	282 ms	14.7 ms	2.0	0	4,050	29.6 s
Sustained decode	64K	4	Valid	179.3	46.9	409 ms	21.3 ms	4.0	0	5,378	30.0 s
Sustained decode	64K	8	Valid	208.3	32.2	792 ms	31.1 ms	8.0	0	6,245	30.0 s
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	120K	1	Valid	64.2	64.4	310 ms	15.5 ms	1.0	0	1,924	30.0 s
Sustained decode	120K	2	Valid	89.3	49.8	506 ms	20.1 ms	2.0	0	2,677	30.0 s
Sustained decode	120K	4	Valid	110.0	32.9	660 ms	30.4 ms	4.0	0	3,298	30.0 s
Sustained decode	120K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	128	Over capacity	—	—	—	—	—	—	—	—

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