Speed test matrix

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

vllm · 0.32.0 · BF165 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, 08:03 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 131072 --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 context10.1K792 ms TTFT

16K context8,6141,842 ms TTFT

32K context6,5244,843 ms TTFT

64K context4,44414,184 ms TTFT

120K context2,83641,622 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	75.023 ms TTFT	149.625 ms TTFT	294.239 ms TTFT	563.844 ms TTFT	1,044.460 ms TTFT	1,770.889 ms TTFT	Over capacity	Over capacity
16K	65.193 ms TTFT	129.8116 ms TTFT	252.5165 ms TTFT	463.6241 ms TTFT	795.2410 ms TTFT	Over capacity	Over capacity	Over capacity
32K	61.7139 ms TTFT	123.3190 ms TTFT	236.2277 ms TTFT	413.4440 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	56.3298 ms TTFT	112.2409 ms TTFT	208.6580 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
120K	48.8602 ms TTFT	96.9818 ms TTFT	Over capacity	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	10.1K	—	792 ms	—	—	1	—	—
Prefill	16K	1	Valid	8,614.0	—	1,842 ms	—	—	1	—	—
Prefill	32K	1	Valid	6,524.0	—	4,843 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,444.0	—	14,184 ms	—	—	1	—	—
Prefill	120K	1	Valid	2,836.0	—	41,622 ms	—	—	1	—	—
Sustained decode	0	1	Valid	75.0	75.3	23 ms	13.3 ms	1.0	0	2,249	30.0 s
Sustained decode	0	2	Valid	149.6	75.1	25 ms	13.3 ms	2.0	0	4,486	30.0 s
Sustained decode	0	4	Valid	294.2	74.0	39 ms	13.5 ms	4.0	0	8,824	30.0 s
Sustained decode	0	8	Valid	563.8	71.0	44 ms	14.1 ms	8.0	0	16,912	30.0 s
Sustained decode	0	16	Valid	1,044.4	66.2	60 ms	15.1 ms	16.0	0	31,335	30.0 s
Sustained decode	0	32	Valid	1,770.8	56.5	89 ms	17.7 ms	32.0	0	53,120	30.0 s
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	65.1	65.2	93 ms	15.3 ms	1.0	0	1,953	30.0 s
Sustained decode	16K	2	Valid	129.8	64.9	116 ms	15.4 ms	2.0	0	3,892	30.0 s
Sustained decode	16K	4	Valid	252.5	63.2	165 ms	15.8 ms	4.0	0	7,576	30.0 s
Sustained decode	16K	8	Valid	463.6	58.1	241 ms	17.2 ms	8.0	0	13,904	30.0 s
Sustained decode	16K	16	Valid	795.2	49.8	410 ms	20.1 ms	16.0	0	23,856	30.0 s
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	61.7	61.8	139 ms	16.2 ms	1.0	0	1,852	30.0 s
Sustained decode	32K	2	Valid	123.3	61.7	190 ms	16.2 ms	2.0	0	3,698	30.0 s
Sustained decode	32K	4	Valid	236.2	59.1	277 ms	16.9 ms	4.0	0	7,084	30.0 s
Sustained decode	32K	8	Valid	413.4	51.7	440 ms	19.3 ms	8.0	0	12,400	30.0 s
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	298 ms	17.7 ms	1.0	0	1,688	30.0 s
Sustained decode	64K	2	Valid	112.2	56.3	409 ms	17.8 ms	2.0	0	3,367	30.0 s
Sustained decode	64K	4	Valid	208.6	52.2	580 ms	19.2 ms	4.0	0	6,256	30.0 s
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	48.8	455.6	602 ms	2.2 ms	1.0	0	1,463	30.0 s
Sustained decode	120K	2	Valid	96.9	48.6	818 ms	20.6 ms	2.0	0	2,908	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	—	—	—	—	—	—	—	—

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