BaşlarkenLiderlik tablosuDecode calculatorModellerOrganizationsReportsMatrix runsDonanımKarşılaştırmalı DeğerlendirmelerPazar yeriKiralıklarProAPI Dokümanlar
Dil
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 0.846011 --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 context14K584 ms TTFT
16K context14.9K1,090 ms TTFT
32K context15K2,153 ms TTFT
64K context15K4,314 ms TTFT
128K context14.3K9,052 ms TTFT
256K context12.6K20,482 ms TTFT
512K context10.1K50,933 ms TTFT
1019K context7,159143,365 ms TTFT
1024K context7,106145,136 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K120.159 ms TTFT220.949 ms TTFT378.2105 ms TTFT602.2145 ms TTFT909.2210 ms TTFT1,345.3367 ms TTFT1,968.3499 ms TTFT2,811.9840 ms TTFT
16K119.595 ms TTFT220.2176 ms TTFT375.1201 ms TTFT604.9315 ms TTFT899.5405 ms TTFT1,332.2640 ms TTFT1,898.91,099 ms TTFT2,777.21,519 ms TTFT
32K119.0149 ms TTFT218.4260 ms TTFT377.4277 ms TTFT596.1358 ms TTFT891.7533 ms TTFT1,294.3853 ms TTFT1,927.11,488 ms TTFT2,752.52,559 ms TTFT
64K118.4209 ms TTFT216.4312 ms TTFT370.1360 ms TTFT594.5549 ms TTFT888.3876 ms TTFT1,299.61,418 ms TTFT1,854.62,699 ms TTFT2,707.45,085 ms TTFT
128K117.5306 ms TTFT214.4460 ms TTFT366.7606 ms TTFT578.9935 ms TTFT862.71,783 ms TTFT1,231.93,278 ms TTFT1,753.26,023 ms TTFT2,595.911,604 ms TTFT
256K115.5602 ms TTFT210.2982 ms TTFT357.51,249 ms TTFT562.42,011 ms TTFT850.43,383 ms TTFT1,211.16,101 ms TTFT1,713.711,808 ms TTFT2,457.423,291 ms TTFT
512K112.11,270 ms TTFT203.71,842 ms TTFT344.52,686 ms TTFT536.74,439 ms TTFT795.97,864 ms TTFT1,111.915,128 ms TTFT1,557.229,747 ms TTFT2,171.862,718 ms TTFT
1019K108.02,580 ms TTFT196.73,817 ms TTFT328.35,460 ms TTFT507.99,278 ms TTFT734.616,539 ms TTFT1,007.532,264 ms TTFT1,305.959,263 ms TTFT1,465.1Requests queued104,729 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
Prefill8K1Valid14K—584 ms——1——
Prefill16K1Valid14.9K—1,090 ms——1——
Prefill32K1Valid15K—2,153 ms——1——
Prefill64K1Valid15K—4,314 ms——1——
Prefill128K1Valid14.3K—9,052 ms——1——
Prefill256K1Valid12.6K—20,482 ms——1——
Prefill512K1Valid10.1K—50,933 ms——1——
Prefill1019K1Valid7,159.0—143,365 ms——1——
Prefill1024K1Valid7,106.0—145,136 ms——1——
Sustained decode8K1Valid120.1120.159 ms8.3 ms1.003,60330.0 s
Sustained decode8K2Valid220.9110.549 ms9.0 ms2.006,62830.0 s
Sustained decode8K4Valid378.294.8105 ms10.5 ms4.0011,34530.0 s
Sustained decode8K8Valid602.275.8145 ms13.2 ms8.0018,06430.0 s
Sustained decode8K16Valid909.257.4210 ms17.4 ms16.0027,26430.0 s
Sustained decode8K32Valid1,345.342.7367 ms23.4 ms32.0040,35330.0 s
Sustained decode8K64Valid1,968.331.3499 ms31.9 ms64.0059,01330.0 s
Sustained decode8K128Valid2,811.922.4840 ms44.7 ms128.0084,35930.0 s
Sustained decode16K1Valid119.5119.595 ms8.4 ms1.003,58430.0 s
Sustained decode16K2Valid220.2110.1176 ms9.1 ms2.006,60730.0 s
Sustained decode16K4Valid375.194.1201 ms10.6 ms4.0011,25230.0 s
Sustained decode16K8Valid604.976.0315 ms13.2 ms8.0018,14430.0 s
Sustained decode16K16Valid899.556.5405 ms17.7 ms16.0026,97730.0 s
Sustained decode16K32Valid1,332.242.1640 ms23.8 ms32.0039,96830.0 s
Sustained decode16K64Valid1,898.930.11,099 ms33.2 ms64.0056,95930.0 s
Sustained decode16K128Valid2,777.221.71,519 ms46.1 ms128.0083,33630.0 s
Sustained decode32K1Valid119.0119.1149 ms8.4 ms1.003,57130.0 s
Sustained decode32K2Valid218.4109.3260 ms9.2 ms2.006,55430.0 s
Sustained decode32K4Valid377.494.7277 ms10.6 ms4.0011,32830.0 s
Sustained decode32K8Valid596.174.8358 ms13.4 ms8.0017,88030.0 s
Sustained decode32K16Valid891.756.1533 ms17.8 ms16.0026,73930.0 s
Sustained decode32K32Valid1,294.340.7853 ms24.6 ms32.0038,81930.0 s
Sustained decode32K64Valid1,927.130.21,488 ms33.1 ms64.0057,79230.0 s
Sustained decode32K128Valid2,752.521.52,559 ms46.6 ms128.0082,56930.0 s
Sustained decode64K1Valid118.4118.6209 ms8.4 ms1.003,55530.0 s
Sustained decode64K2Valid216.4108.3312 ms9.2 ms2.006,49030.0 s
Sustained decode64K4Valid370.192.6360 ms10.8 ms4.0011,10030.0 s
Sustained decode64K8Valid594.574.7549 ms13.4 ms8.0017,83230.0 s
Sustained decode64K16Valid888.355.9876 ms17.9 ms16.0026,63930.0 s
Sustained decode64K32Valid1,299.640.91,418 ms24.5 ms32.0038,97730.0 s
Sustained decode64K64Valid1,854.629.22,699 ms34.3 ms64.0055,61830.0 s
Sustained decode64K128Valid2,707.421.25,085 ms47.1 ms128.0081,15530.0 s
Sustained decode128K1Valid117.5117.7306 ms8.5 ms1.003,52630.0 s
Sustained decode128K2Valid214.4107.5460 ms9.3 ms2.006,43230.0 s
Sustained decode128K4Valid366.791.9606 ms10.9 ms4.0011,00030.0 s
Sustained decode128K8Valid578.972.7935 ms13.8 ms8.0017,36830.0 s
Sustained decode128K16Valid862.754.31,783 ms18.4 ms16.0025,87230.0 s
Sustained decode128K32Valid1,231.939.03,278 ms25.7 ms32.0036,96030.0 s
Sustained decode128K64Valid1,753.228.06,023 ms35.7 ms64.0052,60630.0 s
Sustained decode128K128Valid2,595.920.811,604 ms48.2 ms128.0077,82730.0 s
Sustained decode256K1Valid115.5115.8602 ms8.6 ms1.003,46530.0 s
Sustained decode256K2Valid210.2105.8982 ms9.5 ms2.006,31130.0 s
Sustained decode256K4Valid357.589.61,249 ms11.2 ms4.0010,72430.0 s
Sustained decode256K8Valid562.470.62,011 ms14.2 ms8.0016,87330.0 s
Sustained decode256K16Valid850.453.33,383 ms18.8 ms16.0025,50430.0 s
Sustained decode256K32Valid1,211.138.06,101 ms26.3 ms32.0036,32130.0 s
Sustained decode256K64Valid1,713.727.111,808 ms36.9 ms64.0051,39330.0 s
Sustained decode256K128Valid2,457.419.723,291 ms50.9 ms128.0073,72430.0 s
Sustained decode512K1Valid112.1112.61,270 ms8.9 ms1.003,36430.0 s
Sustained decode512K2Valid203.7102.21,842 ms9.8 ms2.006,11430.0 s
Sustained decode512K4Valid344.586.62,686 ms11.5 ms4.0010,33230.0 s
Sustained decode512K8Valid536.767.54,439 ms14.8 ms8.0016,09630.0 s
Sustained decode512K16Valid795.950.07,864 ms20.0 ms16.0023,87230.0 s
Sustained decode512K32Valid1,111.935.015,128 ms28.6 ms32.0033,34030.0 s
Sustained decode512K64Valid1,557.225.129,747 ms39.8 ms64.0046,72030.0 s
Sustained decode512K128Valid2,171.817.962,718 ms55.8 ms128.0065,15630.0 s
Sustained decode1019K1Valid108.0108.72,580 ms9.2 ms1.003,24130.0 s
Sustained decode1019K2Valid196.798.93,817 ms10.1 ms2.005,90030.0 s
Sustained decode1019K4Valid328.382.85,460 ms12.1 ms4.009,85230.0 s
Sustained decode1019K8Valid507.963.99,278 ms15.6 ms8.0015,23530.0 s
Sustained decode1019K16Valid734.646.316,539 ms21.6 ms16.0022,03330.0 s
Sustained decode1019K32Valid1,007.532.232,264 ms31.0 ms32.0030,21130.0 s
Sustained decode1019K64Valid1,305.921.859,263 ms45.8 ms63.7439,15930.0 s
Sustained decode1019K128Requests queued1,465.117.6104,729 ms56.9 ms91.41243,88730.0 s

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