Speed test matrix

Llama-3.2-3B-Instruct-FP8-dynamic on RTX 3090 24GB

vllm · 0.32.0 · FP8 dynamic5 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 30, 2026, 09:31 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: none
# 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/Llama-3.2-3B-Instruct-FP8-dynamic --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 context14.6K565 ms TTFT

16K context12.1K1,354 ms TTFT

32K context9,0363,595 ms TTFT

64K context5,96810,856 ms TTFT

120K context3,71832,625 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	171.815 ms TTFT	316.219 ms TTFT	565.924 ms TTFT	Looped	1,514.044 ms TTFT	Over capacity	Over capacity	Over capacity
16K	127.957 ms TTFT	202.473 ms TTFT	343.8103 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	101.4104 ms TTFT	146.4137 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	71.1168 ms TTFT	93.7236 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
120K	46.8450 ms TTFT	Over capacity	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	14.6K	—	565 ms	—	—	1	—	—
Prefill	16K	1	Valid	12.1K	—	1,354 ms	—	—	1	—	—
Prefill	32K	1	Valid	9,036.0	—	3,595 ms	—	—	1	—	—
Prefill	64K	1	Valid	5,968.0	—	10,856 ms	—	—	1	—	—
Prefill	120K	1	Valid	3,718.0	—	32,625 ms	—	—	1	—	—
Sustained decode	0	1	Valid	171.8	174.2	15 ms	5.7 ms	1.0	0	5,152	30.0 s
Sustained decode	0	2	Valid	316.2	161.8	19 ms	6.2 ms	2.0	0	9,484	30.0 s
Sustained decode	0	4	Valid	565.9	146.4	24 ms	6.8 ms	4.0	0	16,976	30.0 s
Sustained decode	0	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	16	Valid	1,514.0	102.8	44 ms	9.7 ms	16.0	0	45,424	30.0 s
Sustained decode	0	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	127.9	128.4	57 ms	7.8 ms	1.0	0	3,836	30.0 s
Sustained decode	16K	2	Valid	202.4	101.7	73 ms	9.8 ms	2.0	0	6,070	30.0 s
Sustained decode	16K	4	Valid	343.8	86.6	103 ms	11.5 ms	4.0	0	10,312	30.0 s
Sustained decode	16K	8	Over capacity	—	—	—	—	—	—	—	—
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	101.4	101.7	104 ms	9.8 ms	1.0	0	3,043	30.0 s
Sustained decode	32K	2	Valid	146.4	73.4	137 ms	13.6 ms	2.0	0	4,391	30.0 s
Sustained decode	32K	4	Over capacity	—	—	—	—	—	—	—	—
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	71.1	71.3	168 ms	14.0 ms	1.0	0	2,133	30.0 s
Sustained decode	64K	2	Valid	93.7	47.0	236 ms	21.3 ms	2.0	0	2,810	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	46.8	46.9	450 ms	21.3 ms	1.0	0	1,402	30.0 s
Sustained decode	120K	2	Over capacity	—	—	—	—	—	—	—	—
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.