Speed test matrix

GLM-5.2-W4A16-MTP on CMP 170HX 64GB ×8

vllm · 0.30.1rc1.dev379+gab60b723a · W4A165 contexts (0–96K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 96K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 29, 2026, 11:17 UTC

Hardware

Type: Discrete GPU
GPU: CMP 170HX
GPU count: 8
VRAM: 64 GB (8 × 8 GB)
CPU: Intel Core i7-9800X
RAM: 128 GB
OS: Ubuntu 26.04.1

Engine

Engine: vllm
Version: 0.30.1rc1.dev379+gab60b723a
Quantization: W4A16
Backend: cuda
Repository: github.com/Morrowmake/vllm-cmp170hx
Build: ghcr.io/morrowmake/vllm-cmp170hx@sha256:0d249835cb8502316403350a3cad9d531b46e5249e4076ecb70faa42067b5b7a

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. Warm room, SYSTIN 30-32 C; power: window-average max 174 W, instantaneous peaks up to 332 W (GA100 reports instantaneous draw); enforced limit 180 W
# 8x NVIDIA CMP 170HX 64GB (cmpunlocker), 180 W limit, PCIe Gen2 x4 per card | i7-9800X | 128 GB DDR4 | Ubuntu 26.04.1, NVIDIA driver 610.57.04 | vLLM: Morrowmake vllm-cmp170hx image
vllm serve canada-quant/GLM-5.2-W4A16-MTP --served-model-name bench --trust-remote-code --pipeline-parallel-size 8 --max-model-len 106496 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --host 0.0.0.0 --port 8000 --reasoning-parser glm47 --hf-overrides '{"index_topk_freq":1}'

Prefill

prompt tok/s by context length

8K context1,8024,550 ms TTFT

16K context2,5286,420 ms TTFT

32K context3,11110,387 ms TTFT

64K context3,36919,146 ms TTFT

96K context3,33428,998 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	22.0356 ms TTFT	37.3203 ms TTFT	71.1393 ms TTFT	80.9541 ms TTFT	122.9713 ms TTFT	194.71,054 ms TTFT	297.21,673 ms TTFT	Over capacity
16K	21.1945 ms TTFT	40.9359 ms TTFT	59.5628 ms TTFT	88.9826 ms TTFT	129.51,665 ms TTFT	Over capacity	Over capacity	Over capacity
32K	20.8497 ms TTFT	40.4523 ms TTFT	70.4681 ms TTFT	114.81,283 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
64K	20.7550 ms TTFT	40.3720 ms TTFT	78.2961 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity	Over capacity
96K	20.6731 ms TTFT	40.21,050 ms TTFT	70.91,238 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	1,802.0	—	4,550 ms	—	—	1	—	—
Prefill	16K	1	Valid	2,528.0	—	6,420 ms	—	—	1	—	—
Prefill	32K	1	Valid	3,111.0	—	10,387 ms	—	—	1	—	—
Prefill	64K	1	Valid	3,369.0	—	19,146 ms	—	—	1	—	—
Prefill	96K	1	Valid	3,334.0	—	28,998 ms	—	—	1	—	—
Sustained decode	0	1	Valid	22.0	22.0	356 ms	45.4 ms	1.0	0	660	30.0 s
Sustained decode	0	2	Valid	37.3	18.9	203 ms	52.9 ms	2.0	0	1,119	30.0 s
Sustained decode	0	4	Valid	71.1	17.8	393 ms	56.3 ms	4.0	0	2,134	30.0 s
Sustained decode	0	8	Valid	80.9	10.4	541 ms	96.2 ms	8.0	0	2,425	30.0 s
Sustained decode	0	16	Valid	122.9	8.2	713 ms	121.3 ms	16.0	0	3,681	30.0 s
Sustained decode	0	32	Valid	194.7	6.5	1,054 ms	155.0 ms	32.0	0	5,822	29.9 s
Sustained decode	0	64	Valid	297.2	4.8	1,673 ms	207.4 ms	64.0	0	8,893	29.9 s
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	21.1	21.2	945 ms	47.3 ms	1.0	0	634	30.0 s
Sustained decode	16K	2	Valid	40.9	20.4	359 ms	49.0 ms	2.0	0	1,227	30.0 s
Sustained decode	16K	4	Valid	59.5	14.9	628 ms	67.2 ms	4.0	0	1,784	30.0 s
Sustained decode	16K	8	Valid	88.9	11.1	826 ms	90.2 ms	8.0	0	2,661	29.9 s
Sustained decode	16K	16	Valid	129.5	8.1	1,665 ms	123.0 ms	16.0	0	3,874	29.9 s
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	20.8	20.8	497 ms	48.0 ms	1.0	0	624	30.0 s
Sustained decode	32K	2	Valid	40.4	20.1	523 ms	49.7 ms	2.0	0	1,212	30.0 s
Sustained decode	32K	4	Valid	70.4	17.5	681 ms	57.0 ms	4.0	0	2,113	30.0 s
Sustained decode	32K	8	Valid	114.8	14.3	1,283 ms	70.0 ms	8.0	0	3,440	30.0 s
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	20.7	20.7	550 ms	48.2 ms	1.0	0	621	30.0 s
Sustained decode	64K	2	Valid	40.3	20.2	720 ms	49.6 ms	2.0	0	1,210	30.0 s
Sustained decode	64K	4	Valid	78.2	19.5	961 ms	51.4 ms	4.0	0	2,347	30.0 s
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	Valid	20.6	20.7	731 ms	48.4 ms	1.0	0	619	30.0 s
Sustained decode	96K	2	Valid	40.2	20.1	1,050 ms	49.7 ms	2.0	0	1,205	30.0 s
Sustained decode	96K	4	Valid	70.9	17.6	1,238 ms	56.8 ms	4.0	0	2,126	30.0 s
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	—	—	—	—	—	—	—	—

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