Speed test matrix

TwIL-8B on RTX 3090 24GB

vllm · 0.32.0 · BF165 contexts (0–96K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 96K

by @Sandbergllm-inference-bench v0.8.0Measured Nov 5, 2026, 13:49 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: BF16
Backend: cuda
Repository: github.com/vllm-project/vllm
Build: vllm/vllm-openai:v0.32.0-x86_64

Run

Served model: bench
Revision: 58d49f0d9909e62f9c2fb09e99fc5c76f3f2d630
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 webAI-Official/TwIL-8B --revision 58d49f0d9909e62f9c2fb09e99fc5c76f3f2d630 --max-model-len 106496 --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 context3,2342,535 ms TTFT

16K context2,1987,374 ms TTFT

32K context1,36023,729 ms TTFT

64K context782.082,341 ms TTFT

96K context545.0177,125 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	43.836 ms TTFT	83.849 ms TTFT	162.555 ms TTFT	311.159 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
16K	34.798 ms TTFT	63.4145 ms TTFT	101.3232 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	28.9167 ms TTFT	50.9245 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	21.7301 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
96K	17.3Requests queued449 ms TTFT	Over capacity	Over capacity	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	3,234.0	—	2,535 ms	—	—	1	—	—
Prefill	16K	1	Valid	2,198.0	—	7,374 ms	—	—	1	—	—
Prefill	32K	1	Valid	1,360.0	—	23,729 ms	—	—	1	—	—
Prefill	64K	1	Valid	782.0	—	82,341 ms	—	—	1	—	—
Prefill	96K	1	Valid	545.0	—	177,125 ms	—	—	1	—	—
Sustained decode	0	1	Valid	43.8	43.9	36 ms	22.8 ms	1.0	0	1,315	30.0 s
Sustained decode	0	2	Valid	83.8	42.0	49 ms	23.8 ms	2.0	0	2,512	30.0 s
Sustained decode	0	4	Valid	162.5	40.8	55 ms	24.5 ms	4.0	0	4,876	30.0 s
Sustained decode	0	8	Valid	311.1	39.2	59 ms	25.5 ms	8.0	0	9,328	30.0 s
Sustained decode	0	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	34.7	34.8	98 ms	28.8 ms	1.0	0	1,042	30.0 s
Sustained decode	16K	2	Valid	63.4	31.7	145 ms	31.5 ms	2.0	0	1,902	30.0 s
Sustained decode	16K	4	Valid	101.3	25.4	232 ms	39.4 ms	4.0	0	3,036	30.0 s
Sustained decode	16K	8	Over capacity	—	—	—	—	—	—	—	—
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	28.9	28.9	167 ms	34.6 ms	1.0	0	868	30.0 s
Sustained decode	32K	2	Valid	50.9	25.5	245 ms	39.3 ms	2.0	0	1,528	30.0 s
Sustained decode	32K	4	Over capacity	—	—	—	—	—	—	—	—
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	21.7	21.7	301 ms	46.2 ms	1.0	0	649	30.0 s
Sustained decode	64K	2	Over capacity	—	—	—	—	—	—	—	—
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	96K	1	Requests queued	17.3	17.3	449 ms	57.9 ms	1.0	0	517	30.0 s
Sustained decode	96K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	96K	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.