Pierwsze krokiRankingDecode calculatorModeleOrganizationsReportsMatrix runsSprzętBenchmarkiMarketplaceWynajemProDokumentacja API
Język
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 0.846011 --host 0.0.0.0 --port 8000 --api-key <redacted> --kv-cache-dtype fp8 --block-size 256 --tokenizer-mode deepseek_v4

Prefill

prompt tok/s by context length
8K context14.7K556 ms TTFT
16K context15K1,081 ms TTFT
32K context15.2K2,130 ms TTFT
64K context15.1K4,291 ms TTFT
128K context11.6K11,101 ms TTFT
256K context12.2K21,223 ms TTFT
512K context10.1K50,939 ms TTFT
1019K context7,123144,087 ms TTFT
1024K context7,129144,675 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K120.349 ms TTFT220.869 ms TTFT380.5120 ms TTFT606.3177 ms TTFT910.8241 ms TTFT1,351.2331 ms TTFT1,985.1484 ms TTFT2,877.6876 ms TTFT
16K119.5117 ms TTFT219.8141 ms TTFT377.8198 ms TTFT616.4298 ms TTFT902.8641 ms TTFT1,342.5717 ms TTFT1,966.11,076 ms TTFT2,835.61,593 ms TTFT
32K119.0138 ms TTFT218.1242 ms TTFT375.8250 ms TTFT597.5367 ms TTFT918.1513 ms TTFT1,361.3892 ms TTFT1,969.71,395 ms TTFT2,827.62,687 ms TTFT
64K118.3201 ms TTFT216.6328 ms TTFT369.5318 ms TTFT596.5468 ms TTFT921.1793 ms TTFT1,338.81,398 ms TTFT1,946.82,622 ms TTFT2,783.15,116 ms TTFT
128K117.6305 ms TTFT214.4444 ms TTFT370.5650 ms TTFT587.7955 ms TTFT895.21,605 ms TTFT1,302.43,275 ms TTFT1,879.46,184 ms TTFT2,708.911,959 ms TTFT
256K115.8598 ms TTFT211.0919 ms TTFT363.71,336 ms TTFT581.62,001 ms TTFT869.63,396 ms TTFT1,250.56,263 ms TTFT1,851.011,836 ms TTFT2,565.723,093 ms TTFT
512K112.11,380 ms TTFT203.91,984 ms TTFT349.83,165 ms TTFT549.04,302 ms TTFT826.27,927 ms TTFT1,168.315,974 ms TTFT1,652.633,009 ms TTFT2,253.964,664 ms TTFT
1019K108.02,801 ms TTFT198.24,168 ms TTFT335.46,136 ms TTFT524.810,433 ms TTFT785.120,453 ms TTFT1,053.937,007 ms TTFT1,378.4Requests queued50,884 ms TTFT1,476.0Requests queued105,620 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
Prefill8K1Valid14.7K—556 ms——1——
Prefill16K1Valid15K—1,081 ms——1——
Prefill32K1Valid15.2K—2,130 ms——1——
Prefill64K1Valid15.1K—4,291 ms——1——
Prefill128K1Valid11.6K—11,101 ms——1——
Prefill256K1Valid12.2K—21,223 ms——1——
Prefill512K1Valid10.1K—50,939 ms——1——
Prefill1019K1Valid7,123.0—144,087 ms——1——
Prefill1024K1Valid7,129.0—144,675 ms——1——
Sustained decode8K1Valid120.3120.249 ms8.3 ms1.003,60730.0 s
Sustained decode8K2Valid220.8110.469 ms9.1 ms2.006,62230.0 s
Sustained decode8K4Valid380.595.2120 ms10.5 ms4.0011,41230.0 s
Sustained decode8K8Valid606.376.2177 ms13.1 ms8.0018,18430.0 s
Sustained decode8K16Valid910.857.4241 ms17.4 ms16.0027,31330.0 s
Sustained decode8K32Valid1,351.242.8331 ms23.4 ms32.0040,51230.0 s
Sustained decode8K64Valid1,985.131.5484 ms31.7 ms64.0059,52330.0 s
Sustained decode8K128Valid2,877.622.8876 ms43.9 ms128.0086,27630.0 s
Sustained decode16K1Valid119.5119.6117 ms8.4 ms1.003,58630.0 s
Sustained decode16K2Valid219.8110.0141 ms9.1 ms2.006,59830.0 s
Sustained decode16K4Valid377.894.7198 ms10.6 ms4.0011,33130.0 s
Sustained decode16K8Valid616.477.3298 ms12.9 ms8.0018,48830.0 s
Sustained decode16K16Valid902.857.1641 ms17.5 ms16.0027,07230.0 s
Sustained decode16K32Valid1,342.542.5717 ms23.5 ms32.0040,25730.0 s
Sustained decode16K64Valid1,966.131.11,076 ms32.1 ms64.0058,94330.0 s
Sustained decode16K128Valid2,835.622.21,593 ms45.1 ms128.0084,99430.0 s
Sustained decode32K1Valid119.0119.0138 ms8.4 ms1.003,57030.0 s
Sustained decode32K2Valid218.1109.2242 ms9.2 ms2.006,54230.0 s
Sustained decode32K4Valid375.894.2250 ms10.6 ms4.0011,27230.0 s
Sustained decode32K8Valid597.575.1367 ms13.3 ms8.0017,92030.0 s
Sustained decode32K16Valid918.157.5513 ms17.4 ms16.0027,53630.0 s
Sustained decode32K32Valid1,361.342.7892 ms23.4 ms32.0040,83230.0 s
Sustained decode32K64Valid1,969.730.81,395 ms32.5 ms64.0059,07130.0 s
Sustained decode32K128Valid2,827.622.12,687 ms45.3 ms128.0084,86830.0 s
Sustained decode64K1Valid118.3118.4201 ms8.4 ms1.003,55030.0 s
Sustained decode64K2Valid216.6108.4328 ms9.2 ms2.006,49730.0 s
Sustained decode64K4Valid369.592.6318 ms10.8 ms4.0011,08430.0 s
Sustained decode64K8Valid596.574.8468 ms13.4 ms8.0017,89730.0 s
Sustained decode64K16Valid921.157.8793 ms17.3 ms16.0027,63230.0 s
Sustained decode64K32Valid1,338.842.11,398 ms23.8 ms32.0040,16030.0 s
Sustained decode64K64Valid1,946.830.62,622 ms32.7 ms64.0058,36730.0 s
Sustained decode64K128Valid2,783.121.85,116 ms45.9 ms128.0083,45730.0 s
Sustained decode128K1Valid117.6117.8305 ms8.5 ms1.003,52630.0 s
Sustained decode128K2Valid214.4107.3444 ms9.3 ms2.006,43230.0 s
Sustained decode128K4Valid370.592.8650 ms10.8 ms4.0011,11230.0 s
Sustained decode128K8Valid587.773.8955 ms13.5 ms8.0017,62430.0 s
Sustained decode128K16Valid895.256.11,605 ms17.8 ms16.0026,84830.0 s
Sustained decode128K32Valid1,302.441.03,275 ms24.4 ms32.0039,07130.0 s
Sustained decode128K64Valid1,879.429.96,184 ms33.5 ms64.0056,38330.0 s
Sustained decode128K128Valid2,708.921.711,959 ms46.1 ms128.0081,27930.0 s
Sustained decode256K1Valid115.8116.0598 ms8.6 ms1.003,47230.0 s
Sustained decode256K2Valid211.0105.9919 ms9.4 ms2.006,32830.0 s
Sustained decode256K4Valid363.791.01,336 ms11.0 ms4.0010,90830.0 s
Sustained decode256K8Valid581.672.82,001 ms13.7 ms8.0017,44830.0 s
Sustained decode256K16Valid869.654.43,396 ms18.4 ms16.0026,08130.0 s
Sustained decode256K32Valid1,250.539.26,263 ms25.5 ms32.0037,50530.0 s
Sustained decode256K64Valid1,851.029.111,836 ms34.4 ms64.0055,49030.0 s
Sustained decode256K128Valid2,565.720.323,093 ms49.1 ms128.0076,93030.0 s
Sustained decode512K1Valid112.1112.51,380 ms8.9 ms1.003,36230.0 s
Sustained decode512K2Valid203.9102.41,984 ms9.8 ms2.006,11630.0 s
Sustained decode512K4Valid349.887.93,165 ms11.4 ms4.0010,49230.0 s
Sustained decode512K8Valid549.069.04,302 ms14.5 ms8.0016,46430.0 s
Sustained decode512K16Valid826.251.87,927 ms19.3 ms16.0024,78330.0 s
Sustained decode512K32Valid1,168.336.715,974 ms27.3 ms32.0035,04030.0 s
Sustained decode512K64Valid1,652.626.533,009 ms37.7 ms64.0049,53630.0 s
Sustained decode512K128Valid2,253.918.964,664 ms52.9 ms127.9267,53630.0 s
Sustained decode1019K1Valid108.0108.72,801 ms9.2 ms1.003,23930.0 s
Sustained decode1019K2Valid198.299.84,168 ms10.0 ms2.005,94430.0 s
Sustained decode1019K4Valid335.484.56,136 ms11.8 ms4.0010,06030.0 s
Sustained decode1019K8Valid524.866.010,433 ms15.1 ms8.0015,73630.0 s
Sustained decode1019K16Valid785.149.320,453 ms20.3 ms16.0023,55230.0 s
Sustained decode1019K32Valid1,053.933.837,007 ms29.6 ms31.9131,59630.0 s
Sustained decode1019K64Requests queued1,378.422.550,884 ms44.4 ms63.41941,30930.0 s
Sustained decode1019K128Requests queued1,476.019.5105,620 ms51.2 ms83.11544,28230.0 s

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