시작하기리더보드Decode calculator모델OrganizationsReportsMatrix runs하드웨어벤치마크마켓플레이스렌탈ProAPI 문서
언어
Actual Computer — Every computer, one endpoint

Speed test matrix

DeepSeek-V4-Flash-DSpark 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-DSpark --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 context11.3K726 ms TTFT
16K context11.6K1,402 ms TTFT
32K context11.4K2,834 ms TTFT
64K context11.6K5,580 ms TTFT
128K context11.2K11,505 ms TTFT
256K context10.2K25,181 ms TTFT
512K context8,53160,458 ms TTFT
1019K context6,296163,008 ms TTFT
1024K context6,256164,864 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K223.592 ms TTFT373.8119 ms TTFT854.7143 ms TTFT1,276.9157 ms TTFT1,540.5336 ms TTFT2,447.8614 ms TTFT3,747.81,102 ms TTFT4,047.82,023 ms TTFT
16K381.9115 ms TTFT679.3139 ms TTFT583.2151 ms TTFT1,098.4244 ms TTFT1,729.7373 ms TTFT2,487.9660 ms TTFT3,518.11,069 ms TTFT3,784.11,600 ms TTFT
32K383.7126 ms TTFT467.4187 ms TTFT817.3193 ms TTFT1,026.7281 ms TTFT1,799.2439 ms TTFT2,178.3971 ms TTFT3,465.21,479 ms TTFT3,980.72,697 ms TTFT
64K269.9186 ms TTFT574.3274 ms TTFT719.4281 ms TTFT1,168.1331 ms TTFT1,615.2745 ms TTFT2,592.51,499 ms TTFT3,159.02,917 ms TTFT4,201.15,597 ms TTFT
128K222.0366 ms TTFT486.2460 ms TTFT641.0513 ms TTFT957.1911 ms TTFT1,655.01,671 ms TTFT1,945.33,604 ms TTFT3,038.36,866 ms TTFT3,453.413,166 ms TTFT
256K315.7616 ms TTFT527.91,021 ms TTFT660.4820 ms TTFT822.62,077 ms TTFT1,487.63,420 ms TTFT1,981.57,193 ms TTFT2,987.211,942 ms TTFT3,171.828,525 ms TTFT
512K219.01,340 ms TTFT410.81,493 ms TTFT553.91,801 ms TTFT876.92,713 ms TTFT1,264.96,554 ms TTFT1,681.813,538 ms TTFT2,066.626,581 ms TTFT2,250.653,032 ms TTFT
1019K180.52,670 ms TTFT320.2Requests queued2,954 ms TTFT441.24,799 ms TTFT679.68,529 ms TTFT759.215,807 ms TTFT999.630,677 ms TTFT1,108.555,517 ms TTFT1,179.4Requests queued94,031 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
Prefill8K1Valid11.3K—726 ms——1——
Prefill16K1Valid11.6K—1,402 ms——1——
Prefill32K1Valid11.4K—2,834 ms——1——
Prefill64K1Valid11.6K—5,580 ms——1——
Prefill128K1Valid11.2K—11,505 ms——1——
Prefill256K1Valid10.2K—25,181 ms——1——
Prefill512K1Valid8,531.0—60,458 ms——1——
Prefill1019K1Valid6,296.0—163,008 ms——1——
Prefill1024K1Valid6,256.0—164,864 ms——1——
Sustained decode8K1Valid223.5222.392 ms4.5 ms1.016,70230.0 s
Sustained decode8K2Valid373.8146.9119 ms6.8 ms2.0311,21030.0 s
Sustained decode8K4Valid854.7209.7143 ms4.8 ms4.0425,63630.0 s
Sustained decode8K8Valid1,276.9141.9157 ms7.1 ms8.0938,28430.0 s
Sustained decode8K16Valid1,540.581.2336 ms12.3 ms16.0446,21330.0 s
Sustained decode8K32Valid2,447.864.1614 ms15.6 ms32.0173,42230.0 s
Sustained decode8K64Valid3,747.847.81,102 ms20.9 ms64.00112,38530.0 s
Sustained decode8K128Valid4,047.826.22,023 ms38.2 ms128.00121,23730.0 s
Sustained decode16K1Valid381.9363.4115 ms2.9 ms1.0311,45630.0 s
Sustained decode16K2Valid679.3345.2139 ms2.9 ms2.0520,37130.0 s
Sustained decode16K4Valid583.2143.7151 ms7.0 ms4.0417,49530.0 s
Sustained decode16K8Valid1,098.4113.9244 ms8.8 ms8.0432,93330.0 s
Sustained decode16K16Valid1,729.797.7373 ms10.2 ms16.0451,87030.0 s
Sustained decode16K32Valid2,487.969.8660 ms14.3 ms32.0074,64530.0 s
Sustained decode16K64Valid3,518.148.61,069 ms20.6 ms64.00105,36730.0 s
Sustained decode16K128Valid3,784.126.11,600 ms38.3 ms128.00113,48930.0 s
Sustained decode32K1Valid383.7366.7126 ms2.7 ms1.0311,51030.0 s
Sustained decode32K2Valid467.4217.9187 ms4.6 ms2.0314,01530.0 s
Sustained decode32K4Valid817.3194.3193 ms5.2 ms4.0424,50430.0 s
Sustained decode32K8Valid1,026.7119.3281 ms8.4 ms8.0430,78630.0 s
Sustained decode32K16Valid1,799.291.5439 ms10.9 ms16.0753,94930.0 s
Sustained decode32K32Valid2,178.356.3971 ms17.8 ms32.0165,28930.0 s
Sustained decode32K64Valid3,465.245.81,479 ms21.8 ms64.00103,80230.0 s
Sustained decode32K128Valid3,980.726.42,697 ms37.8 ms128.00119,25930.0 s
Sustained decode64K1Valid269.9226.2186 ms4.4 ms1.028,08830.0 s
Sustained decode64K2Valid574.3331.6274 ms3.0 ms2.0317,22730.0 s
Sustained decode64K4Valid719.4151.5281 ms6.6 ms3.9421,56930.0 s
Sustained decode64K8Valid1,168.1140.1331 ms7.1 ms8.0535,04530.0 s
Sustained decode64K16Valid1,615.284.0745 ms11.9 ms16.0448,45930.0 s
Sustained decode64K32Valid2,592.569.01,499 ms14.5 ms32.0277,76030.0 s
Sustained decode64K64Valid3,159.040.32,917 ms24.8 ms64.0094,71830.0 s
Sustained decode64K128Valid4,201.128.35,597 ms35.3 ms128.00126,00730.0 s
Sustained decode128K1Valid222.0223.8366 ms4.5 ms1.016,65930.0 s
Sustained decode128K2Valid486.2239.8460 ms4.2 ms1.9214,58130.0 s
Sustained decode128K4Valid641.0173.7513 ms5.8 ms4.0419,23030.0 s
Sustained decode128K8Valid957.1113.4911 ms8.8 ms8.0328,69730.0 s
Sustained decode128K16Valid1,655.088.51,671 ms11.3 ms15.9549,59630.0 s
Sustained decode128K32Valid1,945.352.93,604 ms18.9 ms32.0158,25929.9 s
Sustained decode128K64Valid3,038.339.76,866 ms25.2 ms64.0291,06930.0 s
Sustained decode128K128Valid3,453.422.513,166 ms44.5 ms128.00103,61130.0 s
Sustained decode256K1Valid315.7353.2616 ms2.8 ms1.029,47030.0 s
Sustained decode256K2Valid527.9227.01,021 ms4.4 ms1.9415,83030.0 s
Sustained decode256K4Valid660.4140.6820 ms7.1 ms3.9419,80930.0 s
Sustained decode256K8Valid822.6100.52,077 ms10.0 ms8.0224,65930.0 s
Sustained decode256K16Valid1,487.688.63,420 ms11.3 ms15.9444,56730.0 s
Sustained decode256K32Valid1,981.550.77,193 ms19.7 ms31.9359,35430.0 s
Sustained decode256K64Valid2,987.240.011,942 ms25.0 ms63.9389,44129.9 s
Sustained decode256K128Valid3,171.823.228,525 ms43.1 ms127.9495,15730.0 s
Sustained decode512K1Valid219.0228.61,340 ms4.4 ms1.016,56730.0 s
Sustained decode512K2Valid410.8242.41,493 ms4.5 ms1.9212,32330.0 s
Sustained decode512K4Valid553.9132.01,801 ms7.6 ms3.9316,61530.0 s
Sustained decode512K8Valid876.996.82,713 ms10.5 ms7.9426,29730.0 s
Sustained decode512K16Valid1,264.965.96,554 ms15.2 ms15.9337,94530.0 s
Sustained decode512K32Valid1,681.842.813,538 ms23.4 ms31.9250,43930.0 s
Sustained decode512K64Valid2,066.630.826,581 ms32.5 ms63.9161,97930.0 s
Sustained decode512K128Valid2,250.618.253,032 ms54.9 ms127.41567,03329.8 s
Sustained decode1019K1Valid180.5199.32,670 ms5.0 ms0.915,41330.0 s
Sustained decode1019K2Requests queued320.2161.52,954 ms6.2 ms1.829,60230.0 s
Sustained decode1019K4Valid441.2106.14,799 ms9.4 ms3.9113,23330.0 s
Sustained decode1019K8Valid679.694.48,529 ms10.6 ms7.9220,29929.9 s
Sustained decode1019K16Valid759.247.815,807 ms20.9 ms15.9122,76330.0 s
Sustained decode1019K32Valid999.632.230,677 ms31.1 ms31.9229,92129.9 s
Sustained decode1019K64Valid1,108.518.855,517 ms53.3 ms63.61133,19529.9 s
Sustained decode1019K128Requests queued1,179.416.694,031 ms60.1 ms73.41435,14629.8 s

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