Speed test matrix

granite-4.2-3b-fp8 on RTX 3090 24GB

vllm · 0.32.0 · FP85 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, 19:54 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
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 ibm-granite/granite-4.2-3b-fp8 --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 context11.3K731 ms TTFT

16K context9,2301,768 ms TTFT

32K context6,6684,869 ms TTFT

64K context4,36714,830 ms TTFT

120K context2,71944,621 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	156.715 ms TTFT	292.720 ms TTFT	542.119 ms TTFT	938.122 ms TTFT	1,567.146 ms TTFT	Over capacity	Over capacity	Over capacity
16K	124.443 ms TTFT	208.956 ms TTFT	Looped	465.3152 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
32K	105.272 ms TTFT	161.993 ms TTFT	255.4149 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	80.0137 ms TTFT	111.6184 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
120K	56.3261 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	11.3K	—	731 ms	—	—	1	—	—
Prefill	16K	1	Valid	9,230.0	—	1,768 ms	—	—	1	—	—
Prefill	32K	1	Valid	6,668.0	—	4,869 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,367.0	—	14,830 ms	—	—	1	—	—
Prefill	120K	1	Valid	2,719.0	—	44,621 ms	—	—	1	—	—
Sustained decode	0	1	Valid	156.7	158.0	15 ms	6.3 ms	1.0	0	4,700	30.0 s
Sustained decode	0	2	Valid	292.7	148.8	20 ms	6.7 ms	2.0	0	8,780	30.0 s
Sustained decode	0	4	Valid	542.1	138.9	19 ms	7.2 ms	4.0	0	16,260	30.0 s
Sustained decode	0	8	Valid	938.1	122.1	22 ms	8.2 ms	8.0	0	28,143	30.0 s
Sustained decode	0	16	Valid	1,567.1	103.8	46 ms	9.6 ms	16.0	0	47,008	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	124.4	124.9	43 ms	8.0 ms	1.0	0	3,732	30.0 s
Sustained decode	16K	2	Valid	208.9	104.9	56 ms	9.5 ms	2.0	0	6,266	30.0 s
Sustained decode	16K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	8	Valid	465.3	58.8	152 ms	17.0 ms	8.0	0	13,960	30.0 s
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	105.2	105.3	72 ms	9.5 ms	1.0	0	3,156	30.0 s
Sustained decode	32K	2	Valid	161.9	81.1	93 ms	12.3 ms	2.0	0	4,854	30.0 s
Sustained decode	32K	4	Valid	255.4	63.9	149 ms	15.6 ms	4.0	0	7,660	30.0 s
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	80.0	80.1	137 ms	12.5 ms	1.0	0	2,400	30.0 s
Sustained decode	64K	2	Valid	111.6	55.8	184 ms	17.9 ms	2.0	0	3,348	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	56.3	56.4	261 ms	17.7 ms	1.0	0	1,689	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.