Начало работыТаблица лидеровDecode calculatorМоделиOrganizationsReportsMatrix runsОборудованиеБенчмаркиМаркетплейсАрендаProДокументация API
Язык
Actual Computer — Every computer, one endpoint

Speed test matrix

DeepSeek-V4.1-Flash on H200 SXM 141GB ×8

vllm · 0.1.1.dev5+ge77daef89 · 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 SXM
GPU count
8
VRAM
1128 GB (8 × 141 GB)
OS
linux

Engine

Engine
vllm
Version
0.1.1.dev5+ge77daef89
Quantization
MXFP4+FP8
Backend
cuda
Build
cuda-13.2
Commit
e77daef89e18

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-v41-vllm-e77daef89/bin/vllm" serve deepseek-ai/DeepSeek-V4.1-Flash --served-model-name deepseek --trust-remote-code --tensor-parallel-size 8 --reasoning-parser deepseek_v41 --enable-auto-tool-choice --tool-call-parser deepseek_v41 --max-model-len 1048576 --gpu-memory-utilization 0.88 --host 0.0.0.0 --port 8000 --api-key <redacted> --dtype bfloat16 --tokenizer-mode deepseek_v41 --mm-encoder-tp-mode data --engram-config '{"cpu_offload":false}' --speculative-config '{"method":"dspark","num_speculative_tokens":5}'

Prefill

prompt tok/s by context length
8K context19.5K420 ms TTFT
16K context20.1K808 ms TTFT
32K context20.1K1,607 ms TTFT
64K context19.6K3,280 ms TTFT
128K context18.5K6,931 ms TTFT
256K context16.4K15,701 ms TTFT
512K context13.1K39,069 ms TTFT
1019K context7,738132,082 ms TTFT
1024K context7,672133,863 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K254.449 ms TTFT438.176 ms TTFT698.968 ms TTFT1,118.591 ms TTFT1,530.6226 ms TTFT2,320.3357 ms TTFT3,016.1676 ms TTFT2,804.41,569 ms TTFT
16K247.272 ms TTFT425.784 ms TTFT686.2114 ms TTFT1,127.1105 ms TTFT1,549.4328 ms TTFT2,334.4587 ms TTFT2,994.61,066 ms TTFT2,793.62,281 ms TTFT
32K239.6110 ms TTFT435.7121 ms TTFT692.8122 ms TTFT1,101.5146 ms TTFT1,463.3535 ms TTFT2,281.2994 ms TTFT2,904.91,907 ms TTFT2,775.43,697 ms TTFT
64K227.2194 ms TTFT410.2213 ms TTFT662.8214 ms TTFT1,087.7270 ms TTFT1,452.61,047 ms TTFT2,231.92,028 ms TTFT2,860.14,016 ms TTFTOver capacity
128K241.0366 ms TTFT414.2392 ms TTFT699.1548 ms TTFT1,055.7615 ms TTFT1,499.02,274 ms TTFT2,138.44,318 ms TTFT2,695.38,188 ms TTFTOver capacity
256K234.1749 ms TTFT403.9788 ms TTFT649.2907 ms TTFT1,016.51,352 ms TTFT1,394.54,930 ms TTFT1,955.59,315 ms TTFTOver capacityOver capacity
512K220.81,712 ms TTFT381.41,803 ms TTFT587.0Requests queued2,638 ms TTFT899.83,111 ms TTFT1,167.19,089 ms TTFTOver capacityOver capacityOver capacity
1019K193.23,939 ms TTFT330.1Requests queued4,138 ms TTFT524.0Requests queued4,098 ms TTFT762.9Requests queued8,621 ms TTFTOver capacityOver capacityOver capacityOver capacity

Requests queued: the server could not run all requests at once, so some waited in a queue. 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.
PhaseContextRequests at onceStatusTotal tok/sPer request tok/sTTFTTime between tokensAverage active requestsCompleted requestsOutput tokensMeasured for
Prefill8K1Valid19.5K—420 ms——1——
Prefill16K1Valid20.1K—808 ms——1——
Prefill32K1Valid20.1K—1,607 ms——1——
Prefill64K1Valid19.6K—3,280 ms——1——
Prefill128K1Valid18.5K—6,931 ms——1——
Prefill256K1Valid16.4K—15,701 ms——1——
Prefill512K1Valid13.1K—39,069 ms——1——
Prefill1019K1Valid7,738.0—132,082 ms——1——
Prefill1024K1Valid7,672.0—133,863 ms——1——
Sustained decode8K1Valid254.4249.449 ms4.0 ms1.027,63130.0 s
Sustained decode8K2Valid438.1218.476 ms4.6 ms2.0213,14130.0 s
Sustained decode8K4Valid698.9172.868 ms5.8 ms4.0420,95930.0 s
Sustained decode8K8Valid1,118.5136.391 ms7.3 ms8.0833,54130.0 s
Sustained decode8K16Valid1,530.694.8226 ms10.6 ms16.0045,90130.0 s
Sustained decode8K32Valid2,320.372.5357 ms13.8 ms32.0069,55530.0 s
Sustained decode8K64Valid3,016.147.7676 ms21.0 ms64.0090,37230.0 s
Sustained decode8K128Valid2,804.422.51,569 ms44.5 ms128.0084,07530.0 s
Sustained decode16K1Valid247.2247.972 ms4.0 ms1.027,41430.0 s
Sustained decode16K2Valid425.7213.884 ms4.7 ms2.0212,77030.0 s
Sustained decode16K4Valid686.2167.3114 ms6.0 ms4.0420,57930.0 s
Sustained decode16K8Valid1,127.1140.5105 ms7.1 ms8.0833,79630.0 s
Sustained decode16K16Valid1,549.496.1328 ms10.4 ms16.0046,44830.0 s
Sustained decode16K32Valid2,334.473.8587 ms13.5 ms32.0069,98530.0 s
Sustained decode16K64Valid2,994.647.11,066 ms21.2 ms64.0089,81730.0 s
Sustained decode16K128Valid2,793.622.42,281 ms44.6 ms128.0083,74230.0 s
Sustained decode32K1Valid239.6240.1110 ms4.2 ms1.017,18730.0 s
Sustained decode32K2Valid435.7220.3121 ms4.5 ms2.0213,07130.0 s
Sustained decode32K4Valid692.8170.9122 ms5.9 ms4.0420,78130.0 s
Sustained decode32K8Valid1,101.5132.6146 ms7.5 ms8.0833,04630.0 s
Sustained decode32K16Valid1,463.392.6535 ms10.8 ms16.0043,89730.0 s
Sustained decode32K32Valid2,281.271.5994 ms14.0 ms32.0068,40830.0 s
Sustained decode32K64Valid2,904.945.71,907 ms21.9 ms64.0087,07230.0 s
Sustained decode32K128Valid2,775.422.63,697 ms44.3 ms128.0083,13330.0 s
Sustained decode64K1Valid227.2229.5194 ms4.4 ms1.016,81630.0 s
Sustained decode64K2Valid410.2207.9213 ms4.8 ms2.0212,30230.0 s
Sustained decode64K4Valid662.8168.4214 ms5.9 ms4.0419,88230.0 s
Sustained decode64K8Valid1,087.7138.4270 ms7.2 ms8.0832,61430.0 s
Sustained decode64K16Valid1,452.689.41,047 ms11.2 ms16.0043,57730.0 s
Sustained decode64K32Valid2,231.970.12,028 ms14.3 ms32.0066,93230.0 s
Sustained decode64K64Valid2,860.146.04,016 ms21.7 ms64.0085,77430.0 s
Sustained decode64K128Over capacity————————
Sustained decode128K1Valid241.0243.0366 ms4.1 ms1.017,22830.0 s
Sustained decode128K2Valid414.2205.7392 ms4.9 ms2.0212,42430.0 s
Sustained decode128K4Valid699.1175.7548 ms5.7 ms3.9420,97230.0 s
Sustained decode128K8Valid1,055.7132.0615 ms7.6 ms7.9831,67230.0 s
Sustained decode128K16Valid1,499.091.02,274 ms11.0 ms16.0044,95130.0 s
Sustained decode128K32Valid2,138.468.04,318 ms14.7 ms32.0064,13730.0 s
Sustained decode128K64Valid2,695.343.88,188 ms22.8 ms64.0080,75630.0 s
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid234.1237.8749 ms4.2 ms1.017,02230.0 s
Sustained decode256K2Valid403.9208.5788 ms4.8 ms1.9212,11530.0 s
Sustained decode256K4Valid649.2166.8907 ms6.0 ms3.9419,47430.0 s
Sustained decode256K8Valid1,016.5127.21,352 ms7.9 ms7.8730,48130.0 s
Sustained decode256K16Valid1,394.588.34,930 ms11.3 ms16.0041,81030.0 s
Sustained decode256K32Valid1,955.562.19,315 ms16.1 ms32.0058,59930.0 s
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid220.8235.81,712 ms4.2 ms0.916,62330.0 s
Sustained decode512K2Valid381.4204.01,803 ms4.9 ms1.9211,44130.0 s
Sustained decode512K4Requests queued587.0155.72,638 ms6.4 ms3.8417,60230.0 s
Sustained decode512K8Valid899.8121.83,111 ms8.2 ms7.7626,99230.0 s
Sustained decode512K16Valid1,167.175.69,089 ms13.2 ms15.7435,01330.0 s
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid193.2222.03,939 ms4.5 ms0.915,79630.0 s
Sustained decode1019K2Requests queued330.1189.54,138 ms5.3 ms1.829,90030.0 s
Sustained decode1019K4Requests queued524.0154.34,098 ms6.5 ms3.5415,57729.7 s
Sustained decode1019K8Requests queued762.9101.98,621 ms9.8 ms7.4622,88230.0 s
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.