Speed test matrix

MiniCPM5-1B-exl3 on CMP 170HX 64GB

vllm · 0.32.0rc1.dev249+g18305948a · EXL3 3.00 bpw5 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 12:59 UTC

Hardware

Type: Discrete GPU
GPU: CMP 170HX
VRAM: 64 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: EXL3 3.00 bpw
Backend: cuda
Repository: github.com/vllm-project/vllm
Build: ghcr.io/morrowmake/vllm-cmp170hx@sha256:5b45c4d6feafefe2eba962375b0a9a0b64fd22b9e05abb4d1966f14a32a2db79

Run

Served model: bench
Revision: 3.00bpw
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 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 turboderp/MiniCPM5-1B-exl3 --revision 3.00bpw --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code --quantization exl3 --load-format exl3

Prefill

prompt tok/s by context length

8K context5,8831,364 ms TTFT

16K context5,5062,886 ms TTFT

32K context5,1886,100 ms TTFT

64K context4,56313,841 ms TTFT

120K context3,68432,104 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	Looped	222.098 ms TTFT	Looped	687.3133 ms TTFT	Looped	1,743.0261 ms TTFT	Looped	3,349.3557 ms TTFT
16K	107.6152 ms TTFT	Looped	Looped	595.5437 ms TTFT	Looped	Looped	1,558.92,156 ms TTFT	Over capacity
32K	104.5264 ms TTFT	199.2336 ms TTFT	Looped	518.6787 ms TTFT	Looped	941.62,303 ms TTFT	Over capacity	Over capacity
64K	99.3416 ms TTFT	Looped	296.3895 ms TTFT	Looped	545.52,251 ms TTFT	636.34,856 ms TTFT	Over capacity	Over capacity
120K	Looped	152.81,178 ms TTFT	228.01,603 ms TTFT	Looped	360.64,708 ms TTFT	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	5,883.0	—	1,364 ms	—	—	1	—	—
Prefill	16K	1	Valid	5,506.0	—	2,886 ms	—	—	1	—	—
Prefill	32K	1	Valid	5,188.0	—	6,100 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,563.0	—	13,841 ms	—	—	1	—	—
Prefill	120K	1	Valid	3,684.0	—	32,104 ms	—	—	1	—	—
Sustained decode	0	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	2	Valid	222.0	111.2	98 ms	9.0 ms	2.0	0	6,660	30.0 s
Sustained decode	0	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	8	Valid	687.3	86.3	133 ms	11.6 ms	8.0	0	20,608	30.0 s
Sustained decode	0	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	32	Valid	1,743.0	54.9	261 ms	18.2 ms	32.0	0	52,288	30.0 s
Sustained decode	0	64	Looped	—	—	—	—	—	0	0	—
Sustained decode	0	128	Valid	3,349.3	26.3	557 ms	38.0 ms	128.0	0	100,480	30.0 s
Sustained decode	16K	1	Valid	107.6	107.9	152 ms	9.3 ms	1.0	0	3,226	30.0 s
Sustained decode	16K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	8	Valid	595.5	74.6	437 ms	13.4 ms	8.0	0	17,864	30.0 s
Sustained decode	16K	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	32	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	64	Valid	1,558.9	23.8	2,156 ms	42.0 ms	64.0	0	46,784	30.0 s
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	104.5	104.8	264 ms	9.5 ms	1.0	0	3,134	30.0 s
Sustained decode	32K	2	Valid	199.2	99.8	336 ms	10.0 ms	2.0	0	5,974	30.0 s
Sustained decode	32K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	8	Valid	518.6	64.8	787 ms	15.4 ms	8.0	0	15,560	30.0 s
Sustained decode	32K	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	32	Valid	941.6	28.7	2,303 ms	34.8 ms	32.0	0	28,256	30.0 s
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	99.3	99.5	416 ms	10.1 ms	1.0	0	2,979	30.0 s
Sustained decode	64K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	4	Valid	296.3	74.2	895 ms	13.5 ms	4.0	0	8,888	30.0 s
Sustained decode	64K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	16	Valid	545.5	33.7	2,251 ms	29.7 ms	16.0	0	16,368	30.0 s
Sustained decode	64K	32	Valid	636.3	19.2	4,856 ms	52.1 ms	32.0	0	19,072	30.0 s
Sustained decode	64K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	120K	2	Valid	152.8	76.7	1,178 ms	13.0 ms	2.0	0	4,584	30.0 s
Sustained decode	120K	4	Valid	228.0	57.1	1,603 ms	17.5 ms	4.0	0	6,840	30.0 s
Sustained decode	120K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	120K	16	Valid	360.6	22.2	4,708 ms	45.1 ms	16.0	0	10,816	30.0 s
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.