Speed test matrix

granite-4.2-3b-nvfp4 on RTX 3090 24GB

vllm · 0.32.0 · NVFP45 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 07:05 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. KV cache: int4_per_token_head
# 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 ibm-granite/granite-4.2-3b-nvfp4 --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --kv-cache-dtype int4_per_token_head --attention-backend TRITON_ATTN

Prefill

prompt tok/s by context length

8K context5,4521,511 ms TTFT

16K context3,4614,714 ms TTFT

32K context2,01716,100 ms TTFT

64K context1,09459,182 ms TTFT

120K context605.0200,484 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	154.136 ms TTFT	298.015 ms TTFT	525.723 ms TTFT	908.428 ms TTFT	1,448.044 ms TTFT	2,481.565 ms TTFT	3,673.7114 ms TTFT	Over capacity
16K	99.378 ms TTFT	174.8114 ms TTFT	250.5163 ms TTFT	321.4222 ms TTFT	397.3347 ms TTFT	541.3585 ms TTFT	Over capacity	Over capacity
32K	72.2148 ms TTFT	121.2225 ms TTFT	160.5292 ms TTFT	188.6456 ms TTFT	221.9693 ms TTFT	Over capacity	Over capacity	Over capacity
64K	46.3287 ms TTFT	74.4391 ms TTFT	92.3569 ms TTFT	102.2846 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
120K	28.4Requests queued565 ms TTFT	44.4746 ms TTFT	52.81,005 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity

Requests queued: the server could not run all requests at once, so some waited in a queue. 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	5,452.0	—	1,511 ms	—	—	1	—	—
Prefill	16K	1	Valid	3,461.0	—	4,714 ms	—	—	1	—	—
Prefill	32K	1	Valid	2,017.0	—	16,100 ms	—	—	1	—	—
Prefill	64K	1	Valid	1,094.0	—	59,182 ms	—	—	1	—	—
Prefill	120K	1	Valid	605.0	—	200,484 ms	—	—	1	—	—
Sustained decode	0	1	Valid	154.1	156.9	36 ms	6.4 ms	1.0	0	4,621	30.0 s
Sustained decode	0	2	Valid	298.0	152.3	15 ms	6.6 ms	2.0	0	8,940	30.0 s
Sustained decode	0	4	Valid	525.7	136.1	23 ms	7.3 ms	4.0	0	15,772	30.0 s
Sustained decode	0	8	Valid	908.4	119.4	28 ms	8.4 ms	8.0	0	27,248	30.0 s
Sustained decode	0	16	Valid	1,448.0	97.3	44 ms	10.3 ms	16.0	0	43,441	30.0 s
Sustained decode	0	32	Valid	2,481.5	84.9	65 ms	11.8 ms	32.0	0	74,432	30.0 s
Sustained decode	0	64	Valid	3,673.7	63.0	114 ms	15.9 ms	64.0	0	110,208	30.0 s
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	99.3	99.8	78 ms	10.0 ms	1.0	0	2,979	30.0 s
Sustained decode	16K	2	Valid	174.8	87.9	114 ms	11.4 ms	2.0	0	5,244	30.0 s
Sustained decode	16K	4	Valid	250.5	63.0	163 ms	15.9 ms	4.0	0	7,516	30.0 s
Sustained decode	16K	8	Valid	321.4	40.4	222 ms	24.8 ms	8.0	0	9,643	30.0 s
Sustained decode	16K	16	Valid	397.3	24.9	347 ms	40.2 ms	16.0	0	11,920	30.0 s
Sustained decode	16K	32	Valid	541.3	17.0	585 ms	58.9 ms	32.0	0	16,224	30.0 s
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	72.2	72.5	148 ms	13.8 ms	1.0	0	2,166	30.0 s
Sustained decode	32K	2	Valid	121.2	60.8	225 ms	16.5 ms	2.0	0	3,634	30.0 s
Sustained decode	32K	4	Valid	160.5	40.2	292 ms	24.9 ms	4.0	0	4,816	30.0 s
Sustained decode	32K	8	Valid	188.6	23.6	456 ms	42.4 ms	8.0	0	5,656	30.0 s
Sustained decode	32K	16	Valid	221.9	13.9	693 ms	72.1 ms	16.0	0	6,656	30.0 s
Sustained decode	32K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	46.3	46.4	287 ms	21.6 ms	1.0	0	1,389	30.0 s
Sustained decode	64K	2	Valid	74.4	37.3	391 ms	26.8 ms	2.0	0	2,231	30.0 s
Sustained decode	64K	4	Valid	92.3	23.1	569 ms	43.3 ms	4.0	0	2,769	30.0 s
Sustained decode	64K	8	Valid	102.2	12.7	846 ms	78.5 ms	8.0	0	3,064	30.0 s
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	Requests queued	28.4	28.4	565 ms	35.2 ms	1.0	0	852	30.0 s
Sustained decode	120K	2	Valid	44.4	22.2	746 ms	45.1 ms	2.0	0	1,330	30.0 s
Sustained decode	120K	4	Valid	52.8	13.1	1,005 ms	76.2 ms	4.0	0	1,584	30.0 s
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	—	—	—	—	—	—	—	—

Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.