Speed test matrix

Phi-4-reasoning-plus-NVFP4 on RTX 3090 24GB

vllm · 0.32.0 · NVFP43 contexts (0–24K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 24K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 30, 2026, 13:51 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: 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), 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 nvidia/Phi-4-reasoning-plus-NVFP4 --max-model-len 32768 --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 context2,0354,147 ms TTFT

16K context1,48511,119 ms TTFT

24K context1,19120,644 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	70.346 ms TTFT	136.238 ms TTFT	252.841 ms TTFT	449.646 ms TTFT	779.674 ms TTFT	Over capacity	Over capacity	Over capacity
16K	49.190 ms TTFT	86.9141 ms TTFT	126.8213 ms TTFT	164.5263 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
24K	42.8125 ms TTFT	73.6180 ms TTFT	101.0278 ms TTFT	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	2,035.0	—	4,147 ms	—	—	1	—	—
Prefill	16K	1	Valid	1,485.0	—	11,119 ms	—	—	1	—	—
Prefill	24K	1	Valid	1,191.0	—	20,644 ms	—	—	1	—	—
Sustained decode	0	1	Valid	70.3	70.7	46 ms	14.1 ms	1.0	0	2,109	30.0 s
Sustained decode	0	2	Valid	136.2	68.7	38 ms	14.6 ms	2.0	0	4,086	30.0 s
Sustained decode	0	4	Valid	252.8	64.1	41 ms	15.6 ms	4.0	0	7,585	30.0 s
Sustained decode	0	8	Valid	449.6	57.5	46 ms	17.4 ms	8.0	0	13,489	30.0 s
Sustained decode	0	16	Valid	779.6	50.6	74 ms	19.8 ms	16.0	0	23,376	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	49.1	49.2	90 ms	20.3 ms	1.0	0	1,473	30.0 s
Sustained decode	16K	2	Valid	86.9	43.6	141 ms	23.0 ms	2.0	0	2,606	30.0 s
Sustained decode	16K	4	Valid	126.8	31.8	213 ms	31.5 ms	4.0	0	3,804	30.0 s
Sustained decode	16K	8	Valid	164.5	20.6	263 ms	48.5 ms	8.0	0	4,928	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	24K	1	Valid	42.8	42.8	125 ms	23.3 ms	1.0	0	1,283	30.0 s
Sustained decode	24K	2	Valid	73.6	36.9	180 ms	27.1 ms	2.0	0	2,207	30.0 s
Sustained decode	24K	4	Valid	101.0	25.3	278 ms	39.6 ms	4.0	0	3,028	30.0 s
Sustained decode	24K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	24K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	24K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	24K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	24K	128	Over capacity	—	—	—	—	—	—	—	—

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