Speed test matrix

Ornith-1.5-9B-NVFP4 on RTX 3090 24GB

vllm · 0.32.0 · NVFP46 contexts (0–248K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 09:12 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 ornith-ai/Ornith-1.5-9B-NVFP4 --max-model-len 262144 --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 context6,7121,221 ms TTFT

16K context6,3362,561 ms TTFT

32K context5,7955,573 ms TTFT

64K context4,91113,119 ms TTFT

128K context3,76234,212 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	123.334 ms TTFT	228.843 ms TTFT	440.580 ms TTFT	839.1133 ms TTFT	1,450.7237 ms TTFT	2,332.7368 ms TTFT	2,997.4714 ms TTFT	Over capacity
16K	118.3114 ms TTFT	219.4162 ms TTFT	410.6254 ms TTFT	Looped	1,126.6813 ms TTFT	1,541.31,635 ms TTFT	Over capacity	Over capacity
32K	114.6254 ms TTFT	211.0308 ms TTFT	381.8460 ms TTFT	657.3724 ms TTFT	893.11,315 ms TTFT	Over capacity	Over capacity	Over capacity
64K	107.4194 ms TTFT	Looped	333.2376 ms TTFT	523.8607 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
128K	95.1669 ms TTFT	168.7839 ms TTFT	264.31,212 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
248K	78.41,294 ms TTFT	132.21,720 ms TTFT	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	6,712.0	—	1,221 ms	—	—	1	—	—
Prefill	16K	1	Valid	6,336.0	—	2,561 ms	—	—	1	—	—
Prefill	32K	1	Valid	5,795.0	—	5,573 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,911.0	—	13,119 ms	—	—	1	—	—
Prefill	128K	1	Valid	3,762.0	—	34,212 ms	—	—	1	—	—
Sustained decode	0	1	Valid	123.3	123.5	34 ms	8.1 ms	1.0	0	3,699	30.0 s
Sustained decode	0	2	Valid	228.8	114.6	43 ms	8.7 ms	2.0	0	6,864	30.0 s
Sustained decode	0	4	Valid	440.5	110.5	80 ms	9.1 ms	4.0	0	13,212	30.0 s
Sustained decode	0	8	Valid	839.1	105.5	133 ms	9.5 ms	8.0	0	25,168	30.0 s
Sustained decode	0	16	Valid	1,450.7	91.4	237 ms	10.9 ms	16.0	0	43,521	30.0 s
Sustained decode	0	32	Valid	2,332.7	73.4	368 ms	13.6 ms	32.0	0	69,985	30.0 s
Sustained decode	0	64	Valid	2,997.4	47.1	714 ms	21.2 ms	64.0	0	89,935	30.0 s
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	118.3	118.5	114 ms	8.4 ms	1.0	0	3,550	30.0 s
Sustained decode	16K	2	Valid	219.4	109.7	162 ms	9.1 ms	2.0	0	6,582	30.0 s
Sustained decode	16K	4	Valid	410.6	102.5	254 ms	9.8 ms	4.0	0	12,316	30.0 s
Sustained decode	16K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	16	Valid	1,126.6	69.5	813 ms	14.4 ms	16.0	0	33,792	30.0 s
Sustained decode	16K	32	Valid	1,541.3	46.8	1,635 ms	21.4 ms	32.0	0	46,241	30.0 s
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	114.6	114.7	254 ms	8.7 ms	1.0	0	3,439	30.0 s
Sustained decode	32K	2	Valid	211.0	105.4	308 ms	9.5 ms	2.0	0	6,328	30.0 s
Sustained decode	32K	4	Valid	381.8	95.1	460 ms	10.5 ms	4.0	0	11,452	30.0 s
Sustained decode	32K	8	Valid	657.3	81.2	724 ms	12.3 ms	8.0	0	19,720	30.0 s
Sustained decode	32K	16	Valid	893.1	54.3	1,315 ms	18.4 ms	16.0	0	26,784	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	107.4	107.5	194 ms	9.3 ms	1.0	0	3,222	30.0 s
Sustained decode	64K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	4	Valid	333.2	83.4	376 ms	12.0 ms	4.0	0	9,996	30.0 s
Sustained decode	64K	8	Valid	523.8	65.5	607 ms	15.3 ms	8.0	0	15,712	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	128K	1	Valid	95.1	95.3	669 ms	10.5 ms	1.0	0	2,853	30.0 s
Sustained decode	128K	2	Valid	168.7	84.1	839 ms	11.9 ms	2.0	0	5,060	30.0 s
Sustained decode	128K	4	Valid	264.3	65.2	1,212 ms	15.3 ms	4.0	0	7,928	30.0 s
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	248K	1	Valid	78.4	78.6	1,294 ms	12.7 ms	1.0	0	2,351	30.0 s
Sustained decode	248K	2	Valid	132.2	65.7	1,720 ms	15.2 ms	2.0	0	3,966	30.0 s
Sustained decode	248K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	248K	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.