Speed test matrix

gemma-4-E4B-it on RTX 3090 24GB

vllm · 0.32.0 · BF165 contexts (0–104K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 104K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 30, 2026, 23:08 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: BF16
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 google/gemma-4-e4b-it --max-model-len 114688 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --language-model-only

Prefill

prompt tok/s by context length

8K context10.2K787 ms TTFT

16K context8,5931,847 ms TTFT

32K context6,5674,811 ms TTFT

64K context4,44214,188 ms TTFT

104K context3,16232,364 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	74.923 ms TTFT	149.427 ms TTFT	294.139 ms TTFT	563.345 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
16K	65.079 ms TTFT	Looped	252.6148 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	61.8151 ms TTFT	123.4192 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	56.3313 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
104K	50.7542 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	10.2K	—	787 ms	—	—	1	—	—
Prefill	16K	1	Valid	8,593.0	—	1,847 ms	—	—	1	—	—
Prefill	32K	1	Valid	6,567.0	—	4,811 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,442.0	—	14,188 ms	—	—	1	—	—
Prefill	104K	1	Valid	3,162.0	—	32,364 ms	—	—	1	—	—
Sustained decode	0	1	Valid	74.9	75.2	23 ms	13.3 ms	1.0	0	2,247	30.0 s
Sustained decode	0	2	Valid	149.4	75.0	27 ms	13.3 ms	2.0	0	4,484	30.0 s
Sustained decode	0	4	Valid	294.1	73.9	39 ms	13.5 ms	4.0	0	8,820	30.0 s
Sustained decode	0	8	Valid	563.3	71.0	45 ms	14.1 ms	8.0	0	16,896	30.0 s
Sustained decode	0	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	65.0	65.1	79 ms	15.4 ms	1.0	0	1,951	30.0 s
Sustained decode	16K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	4	Valid	252.6	63.2	148 ms	15.8 ms	4.0	0	7,576	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	61.8	61.9	151 ms	16.2 ms	1.0	0	1,853	30.0 s
Sustained decode	32K	2	Valid	123.4	61.8	192 ms	16.2 ms	2.0	0	3,700	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	56.3	56.4	313 ms	17.7 ms	1.0	0	1,687	30.0 s
Sustained decode	64K	2	Over capacity	—	—	—	—	—	—	—	—
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	104K	1	Valid	50.7	50.8	542 ms	19.7 ms	1.0	0	1,520	30.0 s
Sustained decode	104K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	104K	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.