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

DeepSeek-V4-Flash-0731 on H200 NVL 141GB ×2

vllm · 0.25.0 · MXFP4+FP88 contexts (8K–1019K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 256K, 512K, 1019K, 1024K
by @keennaylmx dev importMeasured

Hardware

Type
Discrete GPU
GPU
H200 NVL
GPU count
2
VRAM
282 GB (2 × 141 GB)
OS
linux

Engine

Engine
vllm
Version
0.25.0
Quantization
MXFP4+FP8
Backend
cuda
Build
cuda-13.2

Run

Served model
deepseek
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

"$HOME/env_h200-deepseek-vllm/bin/vllm" serve deepseek-ai/DeepSeek-V4-Flash-0731 --served-model-name deepseek --trust-remote-code --tensor-parallel-size 2 --reasoning-parser deepseek_v4 --enable-auto-tool-choice --tool-call-parser deepseek_v4 --max-model-len 1048576 --gpu-memory-utilization 0.700892 --host 0.0.0.0 --port 8000 --api-key <redacted> --kv-cache-dtype fp8 --block-size 256 --tokenizer-mode deepseek_v4 --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

Prefill

prompt tok/s by context length
8K context11.5K712 ms TTFT
16K context11.7K1,395 ms TTFT
32K context11.4K2,831 ms TTFT
64K context11.4K5,646 ms TTFT
128K context11.1K11,602 ms TTFT
256K context10.2K25,230 ms TTFT
512K context8,53460,434 ms TTFT
1019K context6,305162,765 ms TTFT
1024K context6,256164,847 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K329.393 ms TTFT622.6103 ms TTFT559.2140 ms TTFT1,038.4267 ms TTFT1,535.7308 ms TTFT2,121.0758 ms TTFT3,266.61,120 ms TTFT3,440.91,994 ms TTFT
16K417.2116 ms TTFT513.1145 ms TTFT748.9153 ms TTFT1,133.4233 ms TTFT1,694.4376 ms TTFT2,109.6622 ms TTFT3,372.9967 ms TTFT3,818.91,562 ms TTFT
32K401.6141 ms TTFT411.9190 ms TTFT672.7190 ms TTFT1,226.5248 ms TTFT1,703.3451 ms TTFT2,203.7898 ms TTFT3,454.91,520 ms TTFT3,978.42,775 ms TTFT
64K286.6209 ms TTFT647.6269 ms TTFT740.9281 ms TTFT1,198.6308 ms TTFT1,952.4562 ms TTFT2,558.01,458 ms TTFT3,591.12,767 ms TTFT3,806.35,314 ms TTFT
128K196.0359 ms TTFT500.4407 ms TTFT705.4391 ms TTFT1,171.8601 ms TTFT1,341.11,526 ms TTFT2,052.23,024 ms TTFT3,031.95,842 ms TTFT3,791.411,774 ms TTFT
256K259.9635 ms TTFT364.2792 ms TTFT640.7743 ms TTFT1,241.8921 ms TTFT1,638.12,088 ms TTFT2,369.36,033 ms TTFT2,964.711,420 ms TTFT3,262.224,816 ms TTFT
512K406.51,305 ms TTFT436.11,592 ms TTFT540.51,799 ms TTFT1,031.02,226 ms TTFT1,408.35,351 ms TTFT1,517.513,491 ms TTFT2,088.424,601 ms TTFT2,281.448,482 ms TTFT
1019K193.02,650 ms TTFT254.3Requests queued3,079 ms TTFT479.03,688 ms TTFT542.58,114 ms TTFT733.718,023 ms TTFT851.831,135 ms TTFT1,014.557,826 ms TTFT1,007.7Requests queued107,783 ms TTFT

Requests queued: the server could not run all requests at once, so some waited in a queue.

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
Prefill8K1Valid11.5K—712 ms——1——
Prefill16K1Valid11.7K—1,395 ms——1——
Prefill32K1Valid11.4K—2,831 ms——1——
Prefill64K1Valid11.4K—5,646 ms——1——
Prefill128K1Valid11.1K—11,602 ms——1——
Prefill256K1Valid10.2K—25,230 ms——1——
Prefill512K1Valid8,534.0—60,434 ms——1——
Prefill1019K1Valid6,305.0—162,765 ms——1——
Prefill1024K1Valid6,256.0—164,847 ms——1——
Sustained decode8K1Valid329.3421.793 ms2.4 ms1.029,87530.0 s
Sustained decode8K2Valid622.6358.1103 ms2.8 ms2.0418,67630.0 s
Sustained decode8K4Valid559.2130.1140 ms7.7 ms4.0416,77230.0 s
Sustained decode8K8Valid1,038.4124.0267 ms8.5 ms8.0431,13330.0 s
Sustained decode8K16Valid1,535.771.9308 ms13.9 ms15.9646,06130.0 s
Sustained decode8K32Valid2,121.049.9758 ms20.0 ms32.0063,59330.0 s
Sustained decode8K64Valid3,266.639.11,120 ms25.6 ms64.0097,86630.0 s
Sustained decode8K128Valid3,440.922.71,994 ms44.0 ms128.00102,97229.9 s
Sustained decode16K1Valid417.2390.6116 ms2.6 ms1.0312,51730.0 s
Sustained decode16K2Valid513.1255.7145 ms3.9 ms2.0315,38530.0 s
Sustained decode16K4Valid748.9152.3153 ms6.6 ms4.0522,45530.0 s
Sustained decode16K8Valid1,133.4112.2233 ms8.9 ms7.9533,98430.0 s
Sustained decode16K16Valid1,694.485.0376 ms11.8 ms15.9650,79630.0 s
Sustained decode16K32Valid2,109.649.5622 ms20.2 ms32.0263,24030.0 s
Sustained decode16K64Valid3,372.944.3967 ms22.6 ms64.00101,05730.0 s
Sustained decode16K128Valid3,818.924.91,562 ms40.2 ms128.00114,56630.0 s
Sustained decode32K1Valid401.6361.4141 ms2.8 ms1.0312,04730.0 s
Sustained decode32K2Valid411.9217.0190 ms4.6 ms2.0212,35830.0 s
Sustained decode32K4Valid672.7145.7190 ms6.9 ms4.0420,18130.0 s
Sustained decode32K8Valid1,226.5136.9248 ms7.3 ms7.9636,76530.0 s
Sustained decode32K16Valid1,703.397.7451 ms10.2 ms16.0651,05430.0 s
Sustained decode32K32Valid2,203.757.0898 ms17.5 ms32.0266,07330.0 s
Sustained decode32K64Valid3,454.943.31,520 ms23.1 ms64.00103,45429.9 s
Sustained decode32K128Valid3,978.425.42,775 ms39.4 ms128.00119,21330.0 s
Sustained decode64K1Valid286.6298.8209 ms3.3 ms1.028,59430.0 s
Sustained decode64K2Valid647.6347.8269 ms3.0 ms1.9419,42930.0 s
Sustained decode64K4Valid740.9195.2281 ms5.2 ms4.0422,21330.0 s
Sustained decode64K8Valid1,198.6165.0308 ms6.1 ms7.9635,94830.0 s
Sustained decode64K16Valid1,952.4112.9562 ms8.9 ms15.91058,54730.0 s
Sustained decode64K32Valid2,558.071.31,458 ms14.0 ms32.0476,74230.0 s
Sustained decode64K64Valid3,591.145.92,767 ms21.8 ms64.00107,66230.0 s
Sustained decode64K128Valid3,806.326.35,314 ms38.0 ms128.00114,25230.0 s
Sustained decode128K1Valid196.0201.5359 ms5.0 ms1.015,87730.0 s
Sustained decode128K2Valid500.4317.5407 ms3.1 ms2.0315,01130.0 s
Sustained decode128K4Valid705.4180.1391 ms5.8 ms4.0421,15530.0 s
Sustained decode128K8Valid1,171.8148.1601 ms6.8 ms8.0635,12830.0 s
Sustained decode128K16Valid1,341.168.01,526 ms14.7 ms16.0240,22530.0 s
Sustained decode128K32Valid2,052.249.33,024 ms20.3 ms32.0161,46730.0 s
Sustained decode128K64Valid3,031.936.65,842 ms27.4 ms64.0090,93730.0 s
Sustained decode128K128Valid3,791.425.411,774 ms39.4 ms128.00113,74130.0 s
Sustained decode256K1Valid259.9262.9635 ms3.8 ms1.027,79530.0 s
Sustained decode256K2Valid364.2154.8792 ms6.5 ms2.0210,92530.0 s
Sustained decode256K4Valid640.7131.8743 ms7.6 ms3.9419,20030.0 s
Sustained decode256K8Valid1,241.8159.0921 ms6.3 ms7.9737,22430.0 s
Sustained decode256K16Valid1,638.182.52,088 ms12.1 ms15.9749,07830.0 s
Sustained decode256K32Valid2,369.367.26,033 ms14.9 ms31.9270,98730.0 s
Sustained decode256K64Valid2,964.741.311,420 ms24.2 ms63.9688,35129.8 s
Sustained decode256K128Valid3,262.222.524,816 ms44.5 ms127.8797,45429.9 s
Sustained decode512K1Valid406.5433.81,305 ms2.3 ms0.9312,19030.0 s
Sustained decode512K2Valid436.1212.31,592 ms4.7 ms1.9213,07530.0 s
Sustained decode512K4Valid540.5107.61,799 ms9.3 ms3.9216,20930.0 s
Sustained decode512K8Valid1,031.0133.82,226 ms7.5 ms7.8630,91730.0 s
Sustained decode512K16Valid1,408.378.15,351 ms12.8 ms15.7842,21130.0 s
Sustained decode512K32Valid1,517.537.213,491 ms26.9 ms31.9445,50430.0 s
Sustained decode512K64Valid2,088.429.524,601 ms33.9 ms63.7962,65730.0 s
Sustained decode512K128Valid2,281.417.548,482 ms57.3 ms127.43268,17829.9 s
Sustained decode1019K1Valid193.0245.82,650 ms4.4 ms0.915,78930.0 s
Sustained decode1019K2Requests queued254.3148.03,079 ms6.8 ms1.827,62930.0 s
Sustained decode1019K4Valid479.098.03,688 ms10.2 ms3.9214,36230.0 s
Sustained decode1019K8Valid542.564.48,114 ms15.5 ms7.8216,26230.0 s
Sustained decode1019K16Valid733.744.018,023 ms22.7 ms15.9121,97830.0 s
Sustained decode1019K32Valid851.826.531,135 ms37.9 ms31.9225,52030.0 s
Sustained decode1019K64Valid1,014.517.357,826 ms57.7 ms63.8730,39730.0 s
Sustained decode1019K128Requests queued1,007.712.8107,783 ms78.2 ms92.1930,14229.9 s

Requests queued: the server could not run all requests at once, so some waited in a queue.