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, 13:34 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: 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
# 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 openai/gpt-oss-20b --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code

Prefill

prompt tok/s by context length

8K context9,483860 ms TTFT

16K context7,5862,122 ms TTFT

32K context5,3955,931 ms TTFT

64K context3,38018,871 ms TTFT

120K context2,07757,490 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	165.640 ms TTFT	Looped	423.850 ms TTFT	609.656 ms TTFT	909.973 ms TTFT	1,766.593 ms TTFT	Over capacity	Over capacity
16K	145.153 ms TTFT	182.881 ms TTFT	342.7114 ms TTFT	296.7167 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
32K	137.075 ms TTFT	212.0120 ms TTFT	297.2181 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	106.9111 ms TTFT	164.5194 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
120K	81.9184 ms TTFT	125.9266 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. 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	9,483.0	—	860 ms	—	—	1	—	—
Prefill	16K	1	Valid	7,586.0	—	2,122 ms	—	—	1	—	—
Prefill	32K	1	Valid	5,395.0	—	5,931 ms	—	—	1	—	—
Prefill	64K	1	Valid	3,380.0	—	18,871 ms	—	—	1	—	—
Prefill	120K	1	Valid	2,077.0	—	57,490 ms	—	—	1	—	—
Sustained decode	0	1	Valid	165.6	166.8	40 ms	6.0 ms	1.0	0	4,968	30.0 s
Sustained decode	0	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	4	Valid	423.8	106.4	50 ms	9.4 ms	4.0	0	12,713	30.0 s
Sustained decode	0	8	Valid	609.6	83.5	56 ms	12.0 ms	8.0	0	18,287	30.0 s
Sustained decode	0	16	Valid	909.9	67.1	73 ms	14.9 ms	16.0	0	27,287	30.0 s
Sustained decode	0	32	Valid	1,766.5	60.2	93 ms	16.6 ms	32.0	0	52,982	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	145.1	145.5	53 ms	6.9 ms	1.0	0	4,353	30.0 s
Sustained decode	16K	2	Valid	182.8	114.6	81 ms	8.7 ms	2.0	0	5,477	30.0 s
Sustained decode	16K	4	Valid	342.7	86.3	114 ms	11.6 ms	4.0	0	10,279	30.0 s
Sustained decode	16K	8	Valid	296.7	66.5	167 ms	15.0 ms	8.0	0	8,899	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	137.0	130.1	75 ms	7.7 ms	1.0	0	4,094	29.9 s
Sustained decode	32K	2	Valid	212.0	101.3	120 ms	9.9 ms	2.0	0	6,361	30.0 s
Sustained decode	32K	4	Valid	297.2	74.3	181 ms	13.5 ms	4.0	0	8,915	30.0 s
Sustained decode	32K	8	Over capacity	—	—	—	—	—	—	—	—
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	106.9	107.0	111 ms	9.3 ms	1.0	0	3,206	30.0 s
Sustained decode	64K	2	Valid	164.5	84.1	194 ms	11.9 ms	2.0	0	4,933	30.0 s
Sustained decode	64K	4	Over capacity	—	—	—	—	—	—	—	—
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	120K	1	Valid	81.9	82.1	184 ms	12.2 ms	1.0	0	2,458	30.0 s
Sustained decode	120K	2	Valid	125.9	64.3	266 ms	15.5 ms	2.0	0	3,776	30.0 s
Sustained decode	120K	4	Over capacity	—	—	—	—	—	—	—	—
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	—	—	—	—	—	—	—	—

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. Over capacity: not run because this many requests at this context would exceed the server's KV cache.