入門指南排行榜Decode calculator模型OrganizationsReportsMatrix runs硬體基準評測市集租用ProAPI文件
語言
Actual Computer — Every computer, one endpoint

Speed test matrix

Qwen3.8-Flash-Next on H200 SXM 141GB ×4

vllm · 0.1.dev1+g50a061f79 · BF166 contexts (8K–251K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 251K, 256K
by @keennaylmx dev importMeasured

Hardware

Type
Discrete GPU
GPU
H200 SXM
GPU count
4
VRAM
564 GB (4 × 141 GB)
OS
linux

Engine

Engine
vllm
Version
0.1.dev1+g50a061f79
Quantization
BF16
Backend
cuda
Build
cuda-13.2
Commit
50a061f792f3

Run

Served model
qwen
Revision
main
Decode test
Fixed duration
Prefill mode
Measured during decode warm-up
Time per measurement
30 s
Max output
4,096 tokens

Launch command

env VLLM_PLE_MMAP=1 "$HOME/env_h200-qwen-flash-next-vllm-pr-54129/bin/vllm" serve Qwen/Qwen3.8-Flash-Next --served-model-name qwen --tensor-parallel-size 4 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 262144 --gpu-memory-utilization 0.84 --host 0.0.0.0 --port 8000 --api-key <redacted> --dtype bfloat16 --enable-expert-parallel --speculative-config '{"method":"mtp","num_speculative_tokens":1}'

Prefill

prompt tok/s by context length
8K context24.8K331 ms TTFT
16K context27.7K584 ms TTFT
32K context29.5K1,091 ms TTFT
64K context30.3K2,116 ms TTFT
128K context30.1K4,256 ms TTFT
251K context28.4K8,847 ms TTFT
256K context26K9,857 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K157.0359 ms TTFT266.5291 ms TTFT446.7541 ms TTFT709.7585 ms TTFT1,199.01,489 ms TTFT2,025.01,074 ms TTFT3,383.41,633 ms TTFT5,292.72,732 ms TTFT
16K159.7490 ms TTFT268.9308 ms TTFT458.6577 ms TTFT723.9638 ms TTFT1,213.6857 ms TTFT2,054.11,331 ms TTFT3,413.51,864 ms TTFT5,358.83,022 ms TTFT
32K155.5761 ms TTFT269.9348 ms TTFT458.0644 ms TTFT721.3903 ms TTFT1,214.21,300 ms TTFT2,063.11,842 ms TTFT3,431.02,659 ms TTFT5,325.64,289 ms TTFT
64K156.41,323 ms TTFT269.41,419 ms TTFT455.0920 ms TTFT732.81,337 ms TTFT1,205.21,844 ms TTFT2,061.52,781 ms TTFT3,407.14,508 ms TTFT5,268.38,187 ms TTFT
128K157.32,499 ms TTFTLooped452.91,089 ms TTFT729.01,806 ms TTFT1,194.82,780 ms TTFT2,025.24,788 ms TTFT3,327.38,778 ms TTFT5,159.117,683 ms TTFT
251K156.34,999 ms TTFT265.05,254 ms TTFT446.19,626 ms TTFT715.03,099 ms TTFT1,189.75,598 ms TTFT1,984.910,111 ms TTFT3,187.618,760 ms TTFT4,811.636,546 ms TTFT

Looped: the model kept repeating the same text, so the benchmark discarded this measurement.

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.
PhaseContextRequests at onceStatusTotal tok/sPer request tok/sTTFTTime between tokensAverage active requestsCompleted requestsOutput tokensMeasured for
Prefill8K1Valid24.8K—331 ms——1——
Prefill16K1Valid27.7K—584 ms——1——
Prefill32K1Valid29.5K—1,091 ms——1——
Prefill64K1Valid30.3K—2,116 ms——1——
Prefill128K1Valid30.1K—4,256 ms——1——
Prefill251K1Valid28.4K—8,847 ms——1——
Prefill256K1Valid26K—9,857 ms——1——
Sustained decode8K1Valid157.0160.3359 ms6.2 ms1.014,70930.0 s
Sustained decode8K2Valid266.5133.3291 ms7.5 ms2.027,99430.0 s
Sustained decode8K4Valid446.7112.6541 ms8.9 ms4.0013,39630.0 s
Sustained decode8K8Valid709.781.7585 ms12.2 ms8.0021,28430.0 s
Sustained decode8K16Valid1,199.074.31,489 ms13.5 ms16.0035,96730.0 s
Sustained decode8K32Valid2,025.063.51,074 ms15.8 ms32.0060,75230.0 s
Sustained decode8K64Valid3,383.451.61,633 ms19.4 ms64.00101,43230.0 s
Sustained decode8K128Valid5,292.739.42,732 ms25.4 ms128.00158,69130.0 s
Sustained decode16K1Valid159.7162.5490 ms6.2 ms1.014,79030.0 s
Sustained decode16K2Valid268.9134.5308 ms7.4 ms2.028,06530.0 s
Sustained decode16K4Valid458.6115.1577 ms8.7 ms4.0013,75830.0 s
Sustained decode16K8Valid723.990.7638 ms11.0 ms8.0021,71130.0 s
Sustained decode16K16Valid1,213.675.7857 ms13.2 ms16.0036,39130.0 s
Sustained decode16K32Valid2,054.164.21,331 ms15.6 ms32.0061,60330.0 s
Sustained decode16K64Valid3,413.552.01,864 ms19.2 ms64.00102,40230.0 s
Sustained decode16K128Valid5,358.840.73,022 ms24.6 ms128.00160,63130.0 s
Sustained decode32K1Valid155.5159.7761 ms6.3 ms1.014,66430.0 s
Sustained decode32K2Valid269.9136.7348 ms7.3 ms2.028,09730.0 s
Sustained decode32K4Valid458.0114.3644 ms8.7 ms4.0013,73630.0 s
Sustained decode32K8Valid721.389.8903 ms11.1 ms8.0021,63630.0 s
Sustained decode32K16Valid1,214.275.71,300 ms13.2 ms16.0036,41830.0 s
Sustained decode32K32Valid2,063.163.61,842 ms15.7 ms32.0061,85530.0 s
Sustained decode32K64Valid3,431.052.22,659 ms19.2 ms64.00102,93030.0 s
Sustained decode32K128Valid5,325.639.54,289 ms25.3 ms128.00159,69630.0 s
Sustained decode64K1Valid156.4159.21,323 ms6.3 ms1.014,69030.0 s
Sustained decode64K2Valid269.4133.81,419 ms7.5 ms2.028,08130.0 s
Sustained decode64K4Valid455.0113.2920 ms8.8 ms4.0013,64830.0 s
Sustained decode64K8Valid732.889.71,337 ms11.1 ms8.0021,97830.0 s
Sustained decode64K16Valid1,205.274.31,844 ms13.5 ms16.0036,14730.0 s
Sustained decode64K32Valid2,061.562.42,781 ms16.0 ms32.0061,84230.0 s
Sustained decode64K64Valid3,407.149.24,508 ms20.3 ms64.00102,14430.0 s
Sustained decode64K128Valid5,268.335.48,187 ms28.2 ms128.00157,97830.0 s
Sustained decode128K1Valid157.3162.02,499 ms6.2 ms1.014,71830.0 s
Sustained decode128K2Looped—————00—
Sustained decode128K4Valid452.9112.31,089 ms8.9 ms4.0013,58530.0 s
Sustained decode128K8Valid729.088.91,806 ms11.2 ms8.0021,86630.0 s
Sustained decode128K16Valid1,194.871.02,780 ms14.1 ms16.0035,82830.0 s
Sustained decode128K32Valid2,025.258.04,788 ms17.2 ms32.0060,75230.0 s
Sustained decode128K64Valid3,327.344.18,778 ms22.7 ms64.0099,73730.0 s
Sustained decode128K128Valid5,159.130.317,683 ms33.0 ms128.00154,78230.0 s
Sustained decode251K1Valid156.3164.54,999 ms6.1 ms1.014,69030.0 s
Sustained decode251K2Valid265.0139.25,254 ms7.2 ms1.927,95030.0 s
Sustained decode251K4Valid446.1112.59,626 ms8.9 ms4.0013,38330.0 s
Sustained decode251K8Valid715.087.73,099 ms11.4 ms8.0021,44330.0 s
Sustained decode251K16Valid1,189.770.65,598 ms14.2 ms16.0035,68530.0 s
Sustained decode251K32Valid1,984.955.110,111 ms18.2 ms32.0059,51130.0 s
Sustained decode251K64Valid3,187.641.418,760 ms24.2 ms64.0095,59230.0 s
Sustained decode251K128Valid4,811.629.136,546 ms34.3 ms128.00144,19730.0 s

Looped: the model kept repeating the same text, so the benchmark discarded this measurement.