Per iniziareClassificaDecode calculatorModelliOrganizationsReportsMatrix runsHardwareValutazioni comparativeMarketplaceNoleggiProDocumentazione API
Lingua
Actual Computer — Every computer, one endpoint

Speed test matrix

DeepSeek-V4-Flash 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 --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 0.840939 --host 0.0.0.0 --port 8000 --api-key <redacted> --kv-cache-dtype fp8 --block-size 256 --tokenizer-mode deepseek_v4 --speculative_config '{"method":"mtp","num_speculative_tokens":1}'

Prefill

prompt tok/s by context length
8K context14.6K562 ms TTFT
16K context14.8K1,097 ms TTFT
32K context14.7K2,198 ms TTFT
64K context14.7K4,397 ms TTFT
128K context14K9,249 ms TTFT
256K context12.4K20,870 ms TTFT
512K context9,99751,588 ms TTFT
1019K context7,088144,798 ms TTFT
1024K context7,038146,541 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K190.990 ms TTFT320.9192 ms TTFT542.2115 ms TTFT823.0329 ms TTFT1,329.51,169 ms TTFT1,857.5685 ms TTFT2,746.32,633 ms TTFT3,575.41,560 ms TTFT
16K196.6117 ms TTFT333.6140 ms TTFT566.5140 ms TTFT855.3306 ms TTFT1,266.3439 ms TTFT1,810.1636 ms TTFT2,723.4968 ms TTFT3,477.51,597 ms TTFT
32K197.1131 ms TTFT335.3174 ms TTFT555.0201 ms TTFT846.1355 ms TTFT1,249.6547 ms TTFT1,816.0853 ms TTFT2,720.31,540 ms TTFT3,475.12,612 ms TTFT
64K183.4190 ms TTFT327.1247 ms TTFT538.6261 ms TTFT846.6514 ms TTFT1,212.5823 ms TTFT1,804.21,467 ms TTFT2,630.02,698 ms TTFT3,359.75,188 ms TTFT
128K188.0346 ms TTFT327.9484 ms TTFT511.2445 ms TTFT834.11,023 ms TTFT1,231.31,812 ms TTFT1,791.53,031 ms TTFT2,614.55,784 ms TTFT3,393.111,043 ms TTFT
256K162.7607 ms TTFT317.3772 ms TTFT520.0704 ms TTFT807.61,905 ms TTFT1,127.83,334 ms TTFT1,747.86,149 ms TTFT2,478.411,854 ms TTFT3,092.823,183 ms TTFT
512K173.51,247 ms TTFT297.81,340 ms TTFT486.61,337 ms TTFT789.14,249 ms TTFT1,104.67,646 ms TTFT1,518.414,280 ms TTFT2,171.928,387 ms TTFT2,646.352,092 ms TTFT
1019K151.72,658 ms TTFT262.5Requests queued3,532 ms TTFT438.24,855 ms TTFT674.08,917 ms TTFT985.514,493 ms TTFT1,268.127,664 ms TTFT1,689.642,901 ms TTFT1,669.0Requests queued105,464 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
Prefill8K1Valid14.6K—562 ms——1——
Prefill16K1Valid14.8K—1,097 ms——1——
Prefill32K1Valid14.7K—2,198 ms——1——
Prefill64K1Valid14.7K—4,397 ms——1——
Prefill128K1Valid14K—9,249 ms——1——
Prefill256K1Valid12.4K—20,870 ms——1——
Prefill512K1Valid9,997.0—51,588 ms——1——
Prefill1019K1Valid7,088.0—144,798 ms——1——
Prefill1024K1Valid7,038.0—146,541 ms——1——
Sustained decode8K1Valid190.9193.090 ms5.2 ms1.015,72630.0 s
Sustained decode8K2Valid320.9160.2192 ms6.3 ms2.029,62630.0 s
Sustained decode8K4Valid542.2136.4115 ms7.3 ms4.0416,26530.0 s
Sustained decode8K8Valid823.0103.2329 ms9.7 ms8.0024,69030.0 s
Sustained decode8K16Valid1,329.581.21,169 ms12.3 ms16.0039,88530.0 s
Sustained decode8K32Valid1,857.557.9685 ms17.3 ms32.0055,72030.0 s
Sustained decode8K64Valid2,746.342.52,633 ms23.6 ms64.0082,30330.0 s
Sustained decode8K128Valid3,575.427.51,560 ms36.4 ms128.00107,16730.0 s
Sustained decode16K1Valid196.6196.0117 ms5.1 ms1.015,89830.0 s
Sustained decode16K2Valid333.6163.2140 ms6.1 ms2.0210,00630.0 s
Sustained decode16K4Valid566.5139.7140 ms7.2 ms4.0416,99330.0 s
Sustained decode16K8Valid855.3106.9306 ms9.4 ms8.0025,67730.0 s
Sustained decode16K16Valid1,266.379.4439 ms12.6 ms16.0037,97330.0 s
Sustained decode16K32Valid1,810.155.9636 ms17.9 ms32.0054,29330.0 s
Sustained decode16K64Valid2,723.442.5968 ms23.5 ms64.0081,61730.0 s
Sustained decode16K128Valid3,477.527.01,597 ms37.0 ms128.00104,37030.0 s
Sustained decode32K1Valid197.1194.4131 ms5.1 ms1.015,91230.0 s
Sustained decode32K2Valid335.3167.9174 ms6.0 ms2.0210,05630.0 s
Sustained decode32K4Valid555.0134.8201 ms7.4 ms4.0416,64330.0 s
Sustained decode32K8Valid846.1104.9355 ms9.5 ms8.0025,36930.0 s
Sustained decode32K16Valid1,249.678.2547 ms12.8 ms16.0037,46330.0 s
Sustained decode32K32Valid1,816.056.9853 ms17.6 ms32.0054,47630.0 s
Sustained decode32K64Valid2,720.342.81,540 ms23.4 ms64.0081,62630.0 s
Sustained decode32K128Valid3,475.126.92,612 ms37.1 ms128.00104,20530.0 s
Sustained decode64K1Valid183.4184.5190 ms5.4 ms1.015,50030.0 s
Sustained decode64K2Valid327.1164.6247 ms6.1 ms2.029,81030.0 s
Sustained decode64K4Valid538.6132.5261 ms7.6 ms4.0416,15430.0 s
Sustained decode64K8Valid846.6107.5514 ms9.3 ms8.0025,38630.0 s
Sustained decode64K16Valid1,212.576.5823 ms13.1 ms16.0036,36130.0 s
Sustained decode64K32Valid1,804.256.21,467 ms17.8 ms32.0054,09330.0 s
Sustained decode64K64Valid2,630.040.92,698 ms24.5 ms64.0078,91630.0 s
Sustained decode64K128Valid3,359.726.35,188 ms38.0 ms128.00100,75330.0 s
Sustained decode128K1Valid188.0188.4346 ms5.3 ms1.015,63830.0 s
Sustained decode128K2Valid327.9161.2484 ms6.2 ms2.029,83530.0 s
Sustained decode128K4Valid511.2127.6445 ms7.8 ms4.0415,33630.0 s
Sustained decode128K8Valid834.1101.41,023 ms9.9 ms8.0025,02030.0 s
Sustained decode128K16Valid1,231.377.81,812 ms12.8 ms16.0036,91330.0 s
Sustained decode128K32Valid1,791.555.93,031 ms17.9 ms32.0053,74730.0 s
Sustained decode128K64Valid2,614.540.75,784 ms24.6 ms64.0078,44830.0 s
Sustained decode128K128Valid3,393.126.711,043 ms37.4 ms128.00101,69830.0 s
Sustained decode256K1Valid162.7173.2607 ms5.8 ms1.014,88130.0 s
Sustained decode256K2Valid317.3161.3772 ms6.2 ms2.029,51630.0 s
Sustained decode256K4Valid520.0131.4704 ms7.6 ms3.9415,59730.0 s
Sustained decode256K8Valid807.6100.01,905 ms10.0 ms8.0024,22430.0 s
Sustained decode256K16Valid1,127.871.03,334 ms14.1 ms16.0033,82130.0 s
Sustained decode256K32Valid1,747.855.06,149 ms18.2 ms32.0052,42930.0 s
Sustained decode256K64Valid2,478.439.711,854 ms25.2 ms64.0074,33730.0 s
Sustained decode256K128Valid3,092.825.523,183 ms39.2 ms128.0092,75730.0 s
Sustained decode512K1Valid173.5180.71,247 ms5.5 ms1.015,20530.0 s
Sustained decode512K2Valid297.8156.31,340 ms6.4 ms1.928,93130.0 s
Sustained decode512K4Valid486.6123.31,337 ms8.1 ms3.9314,59630.0 s
Sustained decode512K8Valid789.198.44,249 ms10.2 ms8.0023,66530.0 s
Sustained decode512K16Valid1,104.669.67,646 ms14.4 ms16.0033,13930.0 s
Sustained decode512K32Valid1,518.449.214,280 ms20.3 ms32.0045,52230.0 s
Sustained decode512K64Valid2,171.934.728,387 ms28.8 ms64.0165,10430.0 s
Sustained decode512K128Valid2,646.322.352,092 ms44.8 ms127.8679,42530.0 s
Sustained decode1019K1Valid151.7168.82,658 ms5.9 ms0.914,55030.0 s
Sustained decode1019K2Requests queued262.5150.33,532 ms6.7 ms1.827,87330.0 s
Sustained decode1019K4Valid438.2118.34,855 ms8.5 ms3.8313,14230.0 s
Sustained decode1019K8Valid674.088.58,917 ms11.3 ms7.9120,21230.0 s
Sustained decode1019K16Valid985.562.714,493 ms15.9 ms15.8229,56330.0 s
Sustained decode1019K32Valid1,268.141.127,664 ms24.3 ms31.6438,00230.0 s
Sustained decode1019K64Valid1,689.628.042,901 ms35.7 ms63.12450,64830.0 s
Sustained decode1019K128Requests queued1,669.025.4105,464 ms39.4 ms75.13150,05730.0 s

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