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 31, 2026, 19:37 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: 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 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 context13.1K628 ms TTFT

16K context10.9K1,492 ms TTFT

32K context8,1513,985 ms TTFT

64K context5,36512,075 ms TTFT

120K context3,32436,493 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	168.016 ms TTFT	311.118 ms TTFT	Looped	933.623 ms TTFT	1,494.448 ms TTFT	Over capacity	Over capacity	Over capacity
16K	125.447 ms TTFT	199.360 ms TTFT	325.088 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	99.781 ms TTFT	144.5101 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	70.4153 ms TTFT	92.8209 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
120K	46.3291 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	13.1K	—	628 ms	—	—	1	—	—
Prefill	16K	1	Valid	10.9K	—	1,492 ms	—	—	1	—	—
Prefill	32K	1	Valid	8,151.0	—	3,985 ms	—	—	1	—	—
Prefill	64K	1	Valid	5,365.0	—	12,075 ms	—	—	1	—	—
Prefill	120K	1	Valid	3,324.0	—	36,493 ms	—	—	1	—	—
Sustained decode	0	1	Valid	168.0	170.5	16 ms	5.9 ms	1.0	0	5,038	30.0 s
Sustained decode	0	2	Valid	311.1	159.2	18 ms	6.3 ms	2.0	0	9,332	30.0 s
Sustained decode	0	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	8	Valid	933.6	123.7	23 ms	8.1 ms	8.0	0	28,008	30.0 s
Sustained decode	0	16	Valid	1,494.4	101.6	48 ms	9.8 ms	16.0	0	44,833	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	125.4	125.8	47 ms	7.9 ms	1.0	0	3,760	30.0 s
Sustained decode	16K	2	Valid	199.3	100.1	60 ms	10.0 ms	2.0	0	5,978	30.0 s
Sustained decode	16K	4	Valid	325.0	81.9	88 ms	12.2 ms	4.0	0	9,748	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	99.7	100.1	81 ms	10.0 ms	1.0	0	2,992	30.0 s
Sustained decode	32K	2	Valid	144.5	72.4	101 ms	13.8 ms	2.0	0	4,334	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	70.4	70.5	153 ms	14.2 ms	1.0	0	2,111	30.0 s
Sustained decode	64K	2	Valid	92.8	46.5	209 ms	21.5 ms	2.0	0	2,785	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	46.3	46.3	291 ms	21.6 ms	1.0	0	1,388	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.