Speed test matrix

Qwen3-235B-A22B-Instruct-2507-REAP-nvfp4 on CMP 170HX 64GB ×2

vllm · 0.32.0rc1.dev249+g18305948a · NVFP45 contexts (0–88K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 88K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 11:44 UTC

Hardware

Type: Discrete GPU
GPU: CMP 170HX
GPU count: 2
VRAM: 64 GB (2 × 32 GB)
CPU: 2x Intel Xeon E5-2630 v3
RAM: 64 GB
OS: Ubuntu 26.04.1

Engine

Engine: vllm
Version: 0.32.0rc1.dev249+g18305948a
Quantization: NVFP4
Backend: cuda
Repository: github.com/vllm-project/vllm
Build: ghcr.io/morrowmake/vllm-cmp170hx@sha256:5b45c4d6feafefe2eba962375b0a9a0b64fd22b9e05abb4d1966f14a32a2db79

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. Layout: PP2 (stock vLLM flag)
# 2x NVIDIA CMP 170HX 64GB (cmpunlocker), 180 W limit, PCIe Gen2 x4 | 2x Intel Xeon E5-2630 v3 | 64 GB DDR4 | Ubuntu 26.04.1, NVIDIA driver 610.57.04 | vLLM: Morrowmake vllm-cmp170hx image
vllm serve Banana-Bae/Qwen3-235B-A22B-Instruct-2507-REAP-nvfp4 --max-model-len 98304 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --pipeline-parallel-size 2

Prefill

prompt tok/s by context length

8K context2,3573,482 ms TTFT

16K context2,2727,151 ms TTFT

32K context1,87217,276 ms TTFT

64K context1,32148,836 ms TTFT

88K context1,07382,652 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	50.8159 ms TTFT	78.9159 ms TTFT	132.7157 ms TTFT	175.4216 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
16K	44.7213 ms TTFT	77.1243 ms TTFT	113.4377 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
32K	40.7327 ms TTFT	69.9372 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
64K	34.4563 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
88K	30.7793 ms TTFT	Over capacity	Over capacity	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	2,357.0	—	3,482 ms	—	—	1	—	—
Prefill	16K	1	Valid	2,272.0	—	7,151 ms	—	—	1	—	—
Prefill	32K	1	Valid	1,872.0	—	17,276 ms	—	—	1	—	—
Prefill	64K	1	Valid	1,321.0	—	48,836 ms	—	—	1	—	—
Prefill	88K	1	Valid	1,073.0	—	82,652 ms	—	—	1	—	—
Sustained decode	0	1	Valid	50.8	50.9	159 ms	19.7 ms	1.0	0	1,524	30.0 s
Sustained decode	0	2	Valid	78.9	39.9	159 ms	25.0 ms	2.0	0	2,366	30.0 s
Sustained decode	0	4	Valid	132.7	33.6	157 ms	29.8 ms	4.0	0	3,980	30.0 s
Sustained decode	0	8	Valid	175.4	22.7	216 ms	44.1 ms	8.0	0	5,256	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	44.7	44.8	213 ms	22.3 ms	1.0	0	1,339	30.0 s
Sustained decode	16K	2	Valid	77.1	38.6	243 ms	25.9 ms	2.0	0	2,311	30.0 s
Sustained decode	16K	4	Valid	113.4	28.5	377 ms	35.1 ms	4.0	0	3,402	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	40.7	40.8	327 ms	24.5 ms	1.0	0	1,221	30.0 s
Sustained decode	32K	2	Valid	69.9	35.0	372 ms	28.6 ms	2.0	0	2,097	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	34.4	34.5	563 ms	29.0 ms	1.0	0	1,031	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	88K	1	Valid	30.7	30.8	793 ms	32.5 ms	1.0	0	922	30.0 s
Sustained decode	88K	2	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	4	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	88K	128	Over capacity	—	—	—	—	—	—	—	—

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