Pierwsze krokiRankingDecode calculatorModeleOrganizationsReportsMatrix runsSprzętBenchmarkiMarketplaceWynajemProDokumentacja API
Język
Actual Computer — Every computer, one endpoint

Speed test matrix

DeepSeek-V4.1-Flash on RTX PRO 6000 Blackwell Server Edition 96GB ×8

vllm · 0.30.1rc1.dev147+g02ee13be2 · 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
RTX PRO 6000 Blackwell Server Edition
GPU count
8
VRAM
768 GB (8 × 96 GB)
OS
linux

Engine

Engine
vllm
Version
0.30.1rc1.dev147+g02ee13be2
Quantization
MXFP4+FP8
Backend
cuda
Build
cuda-13.2
Commit
02ee13be2107

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_rtxpro6k-deepseek-v41-vllm-pr-56509/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 0.83 --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":true}' --attention-backend FLASHINFER_MLA_SPARSE_DSV41 --linear-backend flashinfer_cutlass --moe-backend b12x

Prefill

prompt tok/s by context length
8K context10.4K789 ms TTFT
16K context10.6K1,524 ms TTFT
32K context10.6K3,042 ms TTFT
64K context10.4K6,188 ms TTFT
128K context9,92212,953 ms TTFT
256K context9,00628,523 ms TTFT
512K context7,60567,534 ms TTFT
1019K context6,027169,580 ms TTFT
1024K context6,005171,022 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K116.681 ms TTFT217.2131 ms TTFT386.1302 ms TTFT581.1315 ms TTFT828.6606 ms TTFT1,278.31,039 ms TTFT1,743.41,204 ms TTFT3,066.12,584 ms TTFT
16K114.893 ms TTFT213.4152 ms TTFT381.0192 ms TTFT569.0430 ms TTFT826.9728 ms TTFT1,270.4884 ms TTFT1,718.51,291 ms TTFT3,041.92,724 ms TTFT
32K113.0121 ms TTFT213.6204 ms TTFT376.4259 ms TTFT569.3576 ms TTFT813.5963 ms TTFT1,254.51,201 ms TTFT1,692.71,684 ms TTFTOver capacity
64K112.6174 ms TTFT210.4223 ms TTFT373.7308 ms TTFT566.5481 ms TTFT810.9837 ms TTFT1,231.41,496 ms TTFTOver capacityOver capacity
128K112.4298 ms TTFT209.6396 ms TTFT372.0585 ms TTFT559.5955 ms TTFT801.61,704 ms TTFTOver capacityOver capacityOver capacity
256K112.8575 ms TTFT208.4835 ms TTFT367.81,247 ms TTFT552.62,141 ms TTFTOver capacityOver capacityOver capacityOver capacity
512K111.01,275 ms TTFT204.91,853 ms TTFT357.02,859 ms TTFT540.44,940 ms TTFTOver capacityOver capacityOver capacityOver capacity
1019K107.22,800 ms TTFT198.23,781 ms TTFT349.86,207 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacity

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
Prefill8K1Valid10.4K—789 ms——1——
Prefill16K1Valid10.6K—1,524 ms——1——
Prefill32K1Valid10.6K—3,042 ms——1——
Prefill64K1Valid10.4K—6,188 ms——1——
Prefill128K1Valid9,922.0—12,953 ms——1——
Prefill256K1Valid9,006.0—28,523 ms——1——
Prefill512K1Valid7,605.0—67,534 ms——1——
Prefill1019K1Valid6,027.0—169,580 ms——1——
Prefill1024K1Valid6,005.0—171,022 ms——1——
Sustained decode8K1Valid116.6116.681 ms8.6 ms1.003,49930.0 s
Sustained decode8K2Valid217.2108.6131 ms9.2 ms2.006,51630.0 s
Sustained decode8K4Valid386.196.6302 ms10.4 ms4.0011,58430.0 s
Sustained decode8K8Valid581.172.6315 ms13.8 ms8.0017,43230.0 s
Sustained decode8K16Valid828.652.0606 ms19.2 ms16.0024,84730.0 s
Sustained decode8K32Valid1,278.340.01,039 ms25.0 ms32.0038,33530.0 s
Sustained decode8K64Valid1,743.427.01,204 ms37.0 ms64.0052,28830.0 s
Sustained decode8K128Valid3,066.123.62,584 ms42.4 ms128.0091,90630.0 s
Sustained decode16K1Valid114.8114.893 ms8.7 ms1.003,44330.0 s
Sustained decode16K2Valid213.4106.7152 ms9.4 ms2.006,40030.0 s
Sustained decode16K4Valid381.095.2192 ms10.5 ms4.0011,42930.0 s
Sustained decode16K8Valid569.071.3430 ms14.0 ms8.0017,06430.0 s
Sustained decode16K16Valid826.951.9728 ms19.3 ms16.0024,80030.0 s
Sustained decode16K32Valid1,270.439.7884 ms25.2 ms32.0038,11230.0 s
Sustained decode16K64Valid1,718.526.71,291 ms37.5 ms64.0051,52030.0 s
Sustained decode16K128Valid3,041.923.32,724 ms43.0 ms128.0091,26330.0 s
Sustained decode32K1Valid113.0113.1121 ms8.8 ms1.003,39130.0 s
Sustained decode32K2Valid213.6106.8204 ms9.4 ms2.006,40730.0 s
Sustained decode32K4Valid376.493.9259 ms10.7 ms4.0011,28830.0 s
Sustained decode32K8Valid569.371.0576 ms14.1 ms8.0017,07230.0 s
Sustained decode32K16Valid813.550.8963 ms19.7 ms16.0024,39930.0 s
Sustained decode32K32Valid1,254.538.81,201 ms25.8 ms32.0037,63230.0 s
Sustained decode32K64Valid1,692.726.11,684 ms38.3 ms64.0050,75330.0 s
Sustained decode32K128Over capacity————————
Sustained decode64K1Valid112.6112.6174 ms8.9 ms1.003,37630.0 s
Sustained decode64K2Valid210.4105.1223 ms9.5 ms2.006,31030.0 s
Sustained decode64K4Valid373.793.3308 ms10.7 ms4.0011,20830.0 s
Sustained decode64K8Valid566.570.7481 ms14.2 ms8.0016,99230.0 s
Sustained decode64K16Valid810.950.4837 ms19.8 ms16.0024,32030.0 s
Sustained decode64K32Valid1,231.438.11,496 ms26.3 ms32.0036,93030.0 s
Sustained decode64K64Over capacity————————
Sustained decode64K128Over capacity————————
Sustained decode128K1Valid112.4112.6298 ms8.9 ms1.003,37130.0 s
Sustained decode128K2Valid209.6104.8396 ms9.5 ms2.006,28630.0 s
Sustained decode128K4Valid372.092.8585 ms10.8 ms4.0011,15630.0 s
Sustained decode128K8Valid559.569.8955 ms14.3 ms8.0016,78530.0 s
Sustained decode128K16Valid801.650.01,704 ms20.0 ms16.0024,04930.0 s
Sustained decode128K32Over capacity————————
Sustained decode128K64Over capacity————————
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid112.8113.1575 ms8.8 ms1.003,38330.0 s
Sustained decode256K2Valid208.4104.3835 ms9.6 ms2.006,25230.0 s
Sustained decode256K4Valid367.891.81,247 ms10.9 ms4.0011,03230.0 s
Sustained decode256K8Valid552.669.02,141 ms14.5 ms8.0016,57630.0 s
Sustained decode256K16Over capacity————————
Sustained decode256K32Over capacity————————
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid111.0111.41,275 ms9.0 ms1.003,32930.0 s
Sustained decode512K2Valid204.9102.61,853 ms9.7 ms2.006,14630.0 s
Sustained decode512K4Valid357.089.32,859 ms11.2 ms4.0010,70730.0 s
Sustained decode512K8Valid540.467.64,940 ms14.8 ms8.0016,20830.0 s
Sustained decode512K16Over capacity————————
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid107.2107.92,800 ms9.3 ms1.003,21730.0 s
Sustained decode1019K2Valid198.299.33,781 ms10.1 ms2.005,94430.0 s
Sustained decode1019K4Valid349.887.76,207 ms11.4 ms4.0010,49330.0 s
Sustained decode1019K8Over capacity————————
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

Over capacity: not run because this many requests at this context would exceed the server's KV cache.