Speed test matrix

granite-4.2-3b-nvfp4 on CMP 170HX 64GB

vllm · 0.32.0rc1.dev249+g18305948a · NVFP45 contexts (0–120K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 120K

by @Sandbergllm-inference-bench v0.8.0Measured Oct 28, 2026, 19:14 UTCEdited Oct 29, 2026, 00:14 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: 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
# 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 ibm-granite/granite-4.2-3b-nvfp4 --max-model-len 131072 --gpu-memory-utilization 0.94 --max-num-seqs 128 --enable-prefix-caching --trust-remote-code

Prefill

prompt tok/s by context length

8K context10.5K779 ms TTFT

16K context8,8421,833 ms TTFT

32K context6,9454,646 ms TTFT

64K context4,74013,582 ms TTFT

120K context3,01739,952 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	202.455 ms TTFT	370.244 ms TTFT	674.351 ms TTFT	1,167.879 ms TTFT	1,827.7103 ms TTFT	2,737.0196 ms TTFT	3,799.2324 ms TTFT	Over capacity
16K	163.5170 ms TTFT	Looped	423.2305 ms TTFT	595.9391 ms TTFT	707.9674 ms TTFT	Over capacity	Over capacity	Over capacity
32K	137.0252 ms TTFT	210.3302 ms TTFT	298.0490 ms TTFT	377.3787 ms TTFT	416.91,383 ms TTFT	Over capacity	Over capacity	Over capacity
64K	102.5375 ms TTFT	142.8559 ms TTFT	182.0844 ms TTFT	212.61,421 ms TTFT	Over capacity	Over capacity	Over capacity	Over capacity
120K	71.3797 ms TTFT	88.21,088 ms TTFT	105.71,564 ms TTFT	Over capacity	Over capacity	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	10.5K	—	779 ms	—	—	1	—	—
Prefill	16K	1	Valid	8,842.0	—	1,833 ms	—	—	1	—	—
Prefill	32K	1	Valid	6,945.0	—	4,646 ms	—	—	1	—	—
Prefill	64K	1	Valid	4,740.0	—	13,582 ms	—	—	1	—	—
Prefill	120K	1	Valid	3,017.0	—	39,952 ms	—	—	1	—	—
Sustained decode	0	1	Valid	202.4	204.8	55 ms	4.9 ms	1.0	0	6,072	30.0 s
Sustained decode	0	2	Valid	370.2	188.6	44 ms	5.3 ms	2.0	0	11,104	30.0 s
Sustained decode	0	4	Valid	674.3	174.1	51 ms	5.7 ms	4.0	0	20,237	30.0 s
Sustained decode	0	8	Valid	1,167.8	153.2	79 ms	6.5 ms	8.0	0	35,034	30.0 s
Sustained decode	0	16	Valid	1,827.7	123.0	103 ms	8.1 ms	16.0	0	54,832	30.0 s
Sustained decode	0	32	Valid	2,737.0	93.6	196 ms	10.7 ms	32.0	0	82,113	30.0 s
Sustained decode	0	64	Valid	3,799.2	64.6	324 ms	15.5 ms	64.0	0	113,984	30.0 s
Sustained decode	0	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	1	Valid	163.5	164.3	170 ms	6.1 ms	1.0	0	4,904	30.0 s
Sustained decode	16K	2	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	4	Valid	423.2	106.7	305 ms	9.4 ms	4.0	0	12,696	30.0 s
Sustained decode	16K	8	Valid	595.9	74.9	391 ms	13.3 ms	8.0	0	17,872	30.0 s
Sustained decode	16K	16	Valid	707.9	44.3	674 ms	22.6 ms	16.0	0	21,225	30.0 s
Sustained decode	16K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	16K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	1	Valid	137.0	137.8	252 ms	7.3 ms	1.0	0	4,110	30.0 s
Sustained decode	32K	2	Valid	210.3	105.8	302 ms	9.5 ms	2.0	0	6,308	30.0 s
Sustained decode	32K	4	Valid	298.0	74.9	490 ms	13.4 ms	4.0	0	8,940	30.0 s
Sustained decode	32K	8	Valid	377.3	47.2	787 ms	21.2 ms	8.0	0	11,321	30.0 s
Sustained decode	32K	16	Valid	416.9	26.0	1,383 ms	38.5 ms	16.0	0	12,496	30.0 s
Sustained decode	32K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	102.5	103.0	375 ms	9.7 ms	1.0	0	3,074	30.0 s
Sustained decode	64K	2	Valid	142.8	71.7	559 ms	13.9 ms	2.0	0	4,282	30.0 s
Sustained decode	64K	4	Valid	182.0	45.6	844 ms	21.9 ms	4.0	0	5,456	30.0 s
Sustained decode	64K	8	Valid	212.6	26.6	1,421 ms	37.7 ms	8.0	0	6,376	30.0 s
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	120K	1	Valid	71.3	71.6	797 ms	14.0 ms	1.0	0	2,139	30.0 s
Sustained decode	120K	2	Valid	88.2	44.3	1,088 ms	22.6 ms	2.0	0	2,646	30.0 s
Sustained decode	120K	4	Valid	105.7	26.5	1,564 ms	37.8 ms	4.0	0	3,168	30.0 s
Sustained decode	120K	8	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	120K	16	Over capacity	—	—	—	—	—	—	—	—
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.