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

Speed test matrix

DeepSeek-V4-Flash-Vision-Exp on H200 SXM 141GB ×4

sglang · 0.5.6.post3.dev9812+gd5a3b7725 · 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 SXM
GPU count
4
VRAM
564 GB (4 × 141 GB)
OS
linux

Engine

Engine
sglang
Version
0.5.6.post3.dev9812+gd5a3b7725
Quantization
MXFP4+FP8
Backend
cuda
Build
cuda-13.2
Commit
d5a3b7725374

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-vision-sglang-pr-37253/bin/sglang" serve --model-path deepseek-ai/DeepSeek-V4-Flash-Vision-Exp --served-model-name deepseek --trust-remote-code --tp 4 --reasoning-parser deepseek-v4 --tool-call-parser deepseekv4 --context-length 1048576 --mem-fraction-static 0.850786 --enable-metrics --host 0.0.0.0 --port 8000 --api-key <redacted> --moe-runner-backend marlin

Prefill

prompt tok/s by context length
8K context21.2K387 ms TTFT
16K context19.6K831 ms TTFT
32K context20.6K1,568 ms TTFT
64K context20.5K3,148 ms TTFT
128K context18.8K6,843 ms TTFT
256K context16K16,101 ms TTFT
512K context12.3K42,072 ms TTFT
1019K context8,130126,173 ms TTFT
1024K context7,924130,088 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8KLoopedLoopedLoopedLoopedLoopedLoopedLoopedLooped
16K136.9242 ms TTFTLoopedLoopedLoopedLoopedLoopedLoopedLooped
32KLoopedLoopedLoopedLoopedLoopedLoopedLoopedLooped
64K135.4469 ms TTFTLoopedLoopedLoopedLoopedLoopedLoopedLooped
128KLoopedLoopedLoopedLoopedLoopedLoopedLoopedOver capacity
256K128.51,363 ms TTFTLoopedLoopedLoopedLoopedLoopedOver capacityOver capacity
512K130.22,518 ms TTFTLoopedLoopedLoopedLoopedOver capacityOver capacityOver capacity
1019KLoopedLoopedLoopedLoopedOver capacityOver capacityOver capacityOver capacity

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. 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
Prefill8K1Valid21.2K—387 ms——1——
Prefill16K1Valid19.6K—831 ms——1——
Prefill32K1Valid20.6K—1,568 ms——1——
Prefill64K1Valid20.5K—3,148 ms——1——
Prefill128K1Valid18.8K—6,843 ms——1——
Prefill256K1Valid16K—16,101 ms——1——
Prefill512K1Valid12.3K—42,072 ms——1——
Prefill1019K1Valid8,130.0—126,173 ms——1——
Prefill1024K1Valid7,924.0—130,088 ms——1——
Sustained decode8K1Looped—————00—
Sustained decode8K2Looped—————00—
Sustained decode8K4Looped—————00—
Sustained decode8K8Looped—————00—
Sustained decode8K16Looped—————00—
Sustained decode8K32Looped—————00—
Sustained decode8K64Looped—————00—
Sustained decode8K128Looped—————00—
Sustained decode16K1Valid136.9138.2242 ms7.2 ms1.014,10830.0 s
Sustained decode16K2Looped—————00—
Sustained decode16K4Looped—————00—
Sustained decode16K8Looped—————00—
Sustained decode16K16Looped—————00—
Sustained decode16K32Looped—————00—
Sustained decode16K64Looped—————00—
Sustained decode16K128Looped—————00—
Sustained decode32K1Looped—————00—
Sustained decode32K2Looped—————00—
Sustained decode32K4Looped—————00—
Sustained decode32K8Looped—————00—
Sustained decode32K16Looped—————00—
Sustained decode32K32Looped—————00—
Sustained decode32K64Looped—————00—
Sustained decode32K128Looped—————00—
Sustained decode64K1Valid135.4137.1469 ms7.3 ms1.014,06030.0 s
Sustained decode64K2Looped—————00—
Sustained decode64K4Looped—————00—
Sustained decode64K8Looped—————00—
Sustained decode64K16Looped—————00—
Sustained decode64K32Looped—————00—
Sustained decode64K64Looped—————00—
Sustained decode64K128Looped—————00—
Sustained decode128K1Looped—————00—
Sustained decode128K2Looped—————00—
Sustained decode128K4Looped—————00—
Sustained decode128K8Looped—————00—
Sustained decode128K16Looped—————00—
Sustained decode128K32Looped—————00—
Sustained decode128K64Looped—————00—
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid128.5134.41,363 ms7.4 ms1.013,85530.0 s
Sustained decode256K2Looped—————00—
Sustained decode256K4Looped—————00—
Sustained decode256K8Looped—————00—
Sustained decode256K16Looped—————00—
Sustained decode256K32Looped—————00—
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid130.2130.22,518 ms7.7 ms0.913,64428.0 s
Sustained decode512K2Looped—————00—
Sustained decode512K4Looped—————00—
Sustained decode512K8Looped—————00—
Sustained decode512K16Looped—————00—
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Looped—————00—
Sustained decode1019K2Looped—————00—
Sustained decode1019K4Looped—————00—
Sustained decode1019K8Looped—————00—
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

Looped: the model kept repeating the same text, so the benchmark discarded this measurement. Over capacity: not run because this many requests at this context would exceed the server's KV cache.