Speed test matrix

gemma-4-26B-A4B-it-NVFP4 on RTX 3090 24GB

vllm · 0.32.0 · NVFP46 contexts (0–208K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 11:00 UTCEdited Oct 29, 2026, 12:41 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: NVFP4
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/gemma-4-26B-A4B-it-NVFP4 --max-model-len 221184 --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 context7,7541,034 ms TTFT

16K context6,4192,473 ms TTFT

32K context4,8796,477 ms TTFT

64K context3,30019,104 ms TTFT

128K context1,99263,221 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	Looped	264.033 ms TTFT	422.335 ms TTFT	650.146 ms TTFT	Looped	Over capacity	Over capacity	Over capacity
16K	Looped	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity
32K	Looped	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	Looped	Looped	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
128K	82.9675 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
208K	Looped	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	7,754.0	—	1,034 ms	—	—	1	—	—
Prefill	16K	1	Valid	6,419.0	—	2,473 ms	—	—	1	—	—
Prefill	32K	1	Valid	4,879.0	—	6,477 ms	—	—	1	—	—
Prefill	64K	1	Valid	3,300.0	—	19,104 ms	—	—	1	—	—
Prefill	128K	1	Valid	1,992.0	—	63,221 ms	—	—	1	—	—
Sustained decode	0	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	2	Valid	264.0	133.5	33 ms	7.5 ms	2.0	0	7,921	30.0 s
Sustained decode	0	4	Valid	422.3	107.6	35 ms	9.3 ms	4.0	0	12,672	30.0 s
Sustained decode	0	8	Valid	650.1	83.7	46 ms	11.9 ms	8.0	0	19,504	30.0 s
Sustained decode	0	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	8	Looped	—	—	—	—	—	0	0	—
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	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	4	Looped	—	—	—	—	—	0	0	—
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	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	2	Looped	—	—	—	—	—	0	0	—
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	128K	1	Valid	82.9	83.2	675 ms	12.0 ms	1.0	0	2,486	30.0 s
Sustained decode	128K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	208K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	208K	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.