Speed test matrix

Qwen3.5-0.8B on RTX PRO 6000 Blackwell Server Edition 96GB

sglang · 0.5.6.post3.dev9160+g1cf2b8c54 · BF166 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 251K, 256K

by @keennaylmx dev importMeasured Sep 22, 2026, 12:09 UTC

Hardware

Type: Discrete GPU
GPU: RTX PRO 6000 Blackwell Server Edition
VRAM: 96 GB
OS: linux

Engine

Engine: sglang
Version: 0.5.6.post3.dev9160+g1cf2b8c54
Quantization: BF16
Backend: cuda
Repository: github.com/sgl-project/sglang
Build: cuda-13.2
Commit: 1cf2b8c54d81

Run

Served model: qwen
Revision: main
Benchmark tool: llm-inference-bench v0.6.2
Decode test: Fixed duration
Prefill mode: Measured during decode warm-up
Time per measurement: 30 s
Max output: 4,096 tokens

Launch command

"$HOME/env_rtxpro6k-qwen-sglang/bin/sglang" serve --model-path Qwen/Qwen3.5-0.8B --served-model-name qwen --trust-remote-code --tp 1 --reasoning-parser qwen3 --tool-call-parser qwen3_coder --context-length 262144 --mem-fraction-static 0.79 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --enable-flashinfer-allreduce-fusion

Prefill

prompt tok/s by context length

8K context87.5K94 ms TTFT

16K context88.8K183 ms TTFT

32K context79.9K404 ms TTFT

64K context67.2K958 ms TTFT

128K context49.9K2,579 ms TTFT

251K context32.6K7,728 ms TTFT

256K context32.1K8,012 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests

Context	1	2	4	8	16	32	64	128
8K	510.841 ms TTFT	1,005.460 ms TTFT	1,890.963 ms TTFT	3,416.167 ms TTFT	Looped	8,658.7131 ms TTFT	10.8K1,052 ms TTFT	11.5K3,523 ms TTFT
16K	487.666 ms TTFT	962.6115 ms TTFT	1,821.6122 ms TTFT	Looped	Looped	7,696.7191 ms TTFT	Looped	Looped
32K	Looped	897.4141 ms TTFT	1,695.8245 ms TTFT	2,969.9137 ms TTFT	4,607.5139 ms TTFT	6,224.62,043 ms TTFT	Looped	Over capacity
64K	393.5251 ms TTFT	781.9262 ms TTFT	Looped	2,495.7268 ms TTFT	3,524.9274 ms TTFT	4,343.94,143 ms TTFT	Over capacity	Over capacity
128K	317.0481 ms TTFT	626.4510 ms TTFT	1,189.2514 ms TTFT	1,851.7524 ms TTFT	2,394.74,075 ms TTFT	Over capacity	Over capacity	Over capacity
251K	231.4916 ms TTFT	459.31,360 ms TTFT	Looped	1,254.04,097 ms TTFT	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	87.5K	—	94 ms	—	—	1	—	—
Prefill	16K	1	Valid	88.8K	—	183 ms	—	—	1	—	—
Prefill	32K	1	Valid	79.9K	—	404 ms	—	—	1	—	—
Prefill	64K	1	Valid	67.2K	—	958 ms	—	—	1	—	—
Prefill	128K	1	Valid	49.9K	—	2,579 ms	—	—	1	—	—
Prefill	251K	1	Valid	32.6K	—	7,728 ms	—	—	1	—	—
Prefill	256K	1	Valid	32.1K	—	8,012 ms	—	—	1	—	—
Sustained decode	8K	1	Valid	510.8	514.0	41 ms	1.9 ms	1.0	4	15,323	30.0 s
Sustained decode	8K	2	Valid	1,005.4	507.0	60 ms	2.0 ms	2.0	8	30,163	30.0 s
Sustained decode	8K	4	Valid	1,890.9	476.4	63 ms	2.1 ms	4.0	12	56,725	30.0 s
Sustained decode	8K	8	Valid	3,416.1	431.0	67 ms	2.3 ms	7.9	24	102,484	30.0 s
Sustained decode	8K	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	8K	32	Valid	8,658.7	277.5	131 ms	3.6 ms	31.7	64	259,760	30.0 s
Sustained decode	8K	64	Valid	10.8K	178.9	1,052 ms	5.6 ms	63.8	64	324,745	30.0 s
Sustained decode	8K	128	Valid	11.5K	96.5	3,523 ms	10.4 ms	128.0	0	344,480	30.0 s
Sustained decode	16K	1	Valid	487.6	491.2	66 ms	2.0 ms	1.0	3	14,628	30.0 s
Sustained decode	16K	2	Valid	962.6	489.1	115 ms	2.0 ms	2.0	8	28,879	30.0 s
Sustained decode	16K	4	Valid	1,821.6	461.9	122 ms	2.2 ms	4.0	12	54,649	30.0 s
Sustained decode	16K	8	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	16	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	32	Valid	7,696.7	245.1	191 ms	4.1 ms	31.7	57	230,902	30.0 s
Sustained decode	16K	64	Looped	—	—	—	—	—	0	0	—
Sustained decode	16K	128	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	1	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	2	Valid	897.4	456.0	141 ms	2.2 ms	2.0	6	26,921	30.0 s
Sustained decode	32K	4	Valid	1,695.8	435.0	245 ms	2.3 ms	4.0	12	50,873	30.0 s
Sustained decode	32K	8	Valid	2,969.9	377.9	137 ms	2.6 ms	7.9	24	89,095	30.0 s
Sustained decode	32K	16	Valid	4,607.5	294.3	139 ms	3.4 ms	15.9	32	138,224	30.0 s
Sustained decode	32K	32	Valid	6,224.6	200.9	2,043 ms	5.0 ms	31.8	32	186,739	30.0 s
Sustained decode	32K	64	Looped	—	—	—	—	—	0	0	—
Sustained decode	32K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	1	Valid	393.5	404.2	251 ms	2.5 ms	1.0	3	11,803	30.0 s
Sustained decode	64K	2	Valid	781.9	401.7	262 ms	2.5 ms	1.9	6	23,457	30.0 s
Sustained decode	64K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	64K	8	Valid	2,495.7	319.3	268 ms	3.1 ms	7.8	16	74,868	30.0 s
Sustained decode	64K	16	Valid	3,524.9	223.8	274 ms	4.5 ms	15.8	20	105,739	30.0 s
Sustained decode	64K	32	Valid	4,343.9	143.9	4,143 ms	6.9 ms	31.6	32	130,320	30.0 s
Sustained decode	64K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	64K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	1	Valid	317.0	327.5	481 ms	3.1 ms	1.0	2	9,511	30.0 s
Sustained decode	128K	2	Valid	626.4	324.9	510 ms	3.1 ms	1.9	4	18,791	30.0 s
Sustained decode	128K	4	Valid	1,189.2	310.6	514 ms	3.2 ms	3.9	8	35,674	30.0 s
Sustained decode	128K	8	Valid	1,851.7	245.0	524 ms	4.1 ms	7.8	13	55,250	29.8 s
Sustained decode	128K	16	Valid	2,394.7	153.9	4,075 ms	6.5 ms	15.8	16	71,834	30.0 s
Sustained decode	128K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	128K	128	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	1	Valid	231.4	238.9	916 ms	4.2 ms	1.0	1	6,942	30.0 s
Sustained decode	251K	2	Valid	459.3	237.6	1,360 ms	4.2 ms	1.9	2	13,777	30.0 s
Sustained decode	251K	4	Looped	—	—	—	—	—	0	0	—
Sustained decode	251K	8	Valid	1,254.0	164.2	4,097 ms	6.1 ms	7.7	8	37,615	30.0 s
Sustained decode	251K	16	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	32	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	64	Over capacity	—	—	—	—	—	—	—	—
Sustained decode	251K	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.