はじめにリーダーボードDecode calculatorモデルOrganizationsReportsMatrix runsハードウェアベンチマークマーケットプレイスレンタルProAPIドキュメント
言語
Actual Computer — Every computer, one endpoint

Speed test matrix

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 on H200 NVL 141GB

vllm · 0.26.1rc1.dev638+g52be12cfa · NVFP4+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
VRAM
141 GB
OS
linux

Engine

Engine
vllm
Version
0.26.1rc1.dev638+g52be12cfa
Quantization
NVFP4+FP8
Backend
cuda
Build
cuda-13.2
Commit
52be12cfac0c

Run

Served model
nemotron
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-nvidia-nemotron/bin/vllm" serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 --served-model-name nemotron --trust-remote-code --tensor-parallel-size 1 --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 1048576 --gpu-memory-utilization 0.886040 --host 0.0.0.0 --port 8000 --api-key <redacted> --max-num-seqs 256 --max-num-batched-tokens 16384 --async-scheduling --mamba-backend flashinfer --mamba-ssu-algorithm horizontal --mamba-cache-mode align --mamba-ssm-cache-dtype float16 --enable-mamba-cache-stochastic-rounding --mamba-cache-philox-rounds 5 --linear-backend humming --moe-backend humming

Prefill

prompt tok/s by context length
8K context41K199 ms TTFT
16K context46.3K350 ms TTFT
32K context46K702 ms TTFT
64K context43.3K1,486 ms TTFT
128K context36.6K3,510 ms TTFT
256K context26.6K9,654 ms TTFT
512K context17K30,159 ms TTFT
1019K context9,920103,078 ms TTFT
1024K context9,908103,714 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K292.381 ms TTFT514.0142 ms TTFT805.3204 ms TTFT1,197.8334 ms TTFT1,696.7495 ms TTFT2,451.7744 ms TTFT3,805.21,305 ms TTFT5,827.82,322 ms TTFT
16K290.5106 ms TTFT510.7158 ms TTFT804.9213 ms TTFT1,183.9345 ms TTFT1,656.6432 ms TTFT2,384.9717 ms TTFT3,681.01,289 ms TTFT5,602.42,034 ms TTFT
32K287.7107 ms TTFT508.2117 ms TTFT779.3193 ms TTFT1,145.8256 ms TTFT1,624.9482 ms TTFT2,341.1651 ms TTFT3,496.91,094 ms TTFT5,112.32,205 ms TTFT
64K283.0166 ms TTFT495.5200 ms TTFT759.8294 ms TTFT1,100.2337 ms TTFT1,545.9895 ms TTFT2,175.91,186 ms TTFT3,142.52,163 ms TTFT4,411.84,118 ms TTFT
128K269.9263 ms TTFT477.9334 ms TTFT722.3341 ms TTFT1,013.0350 ms TTFT1,393.01,333 ms TTFT1,881.82,549 ms TTFT2,584.85,114 ms TTFT3,413.89,586 ms TTFT
256K248.3561 ms TTFT444.0705 ms TTFT633.9874 ms TTFT911.51,683 ms TTFT1,174.72,931 ms TTFT1,517.65,318 ms TTFT1,935.010,372 ms TTFT2,370.121,102 ms TTFT
512K219.31,037 ms TTFT385.21,277 ms TTFT530.41,242 ms TTFT742.03,569 ms TTFT914.76,628 ms TTFT1,093.912,094 ms TTFT1,285.623,207 ms TTFTOver capacity
1019K170.82,386 ms TTFT303.02,613 ms TTFT426.54,818 ms TTFT529.28,023 ms TTFT615.414,312 ms TTFT695.826,748 ms TTFTOver 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
Prefill8K1Valid41K—199 ms——1——
Prefill16K1Valid46.3K—350 ms——1——
Prefill32K1Valid46K—702 ms——1——
Prefill64K1Valid43.3K—1,486 ms——1——
Prefill128K1Valid36.6K—3,510 ms——1——
Prefill256K1Valid26.6K—9,654 ms——1——
Prefill512K1Valid17K—30,159 ms——1——
Prefill1019K1Valid9,920.0—103,078 ms——1——
Prefill1024K1Valid9,908.0—103,714 ms——1——
Sustained decode8K1Valid292.3294.081 ms3.4 ms1.028,77030.0 s
Sustained decode8K2Valid514.0259.5142 ms3.9 ms2.0415,42830.0 s
Sustained decode8K4Valid805.3203.0204 ms4.9 ms4.0424,16030.0 s
Sustained decode8K8Valid1,197.8152.5334 ms6.6 ms8.0835,93130.0 s
Sustained decode8K16Valid1,696.7107.3495 ms9.3 ms16.0050,89730.0 s
Sustained decode8K32Valid2,451.777.3744 ms12.9 ms32.0073,53530.0 s
Sustained decode8K64Valid3,805.259.51,305 ms16.8 ms64.00114,17830.0 s
Sustained decode8K128Valid5,827.844.32,322 ms22.6 ms128.00174,79830.0 s
Sustained decode16K1Valid290.5292.7106 ms3.4 ms1.028,71330.0 s
Sustained decode16K2Valid510.7258.3158 ms3.9 ms2.0415,32030.0 s
Sustained decode16K4Valid804.9202.9213 ms4.9 ms4.0424,14430.0 s
Sustained decode16K8Valid1,183.9150.2345 ms6.7 ms8.0835,51830.0 s
Sustained decode16K16Valid1,656.6104.8432 ms9.5 ms16.0049,69830.0 s
Sustained decode16K32Valid2,384.975.3717 ms13.3 ms32.0071,55130.0 s
Sustained decode16K64Valid3,681.058.01,289 ms17.2 ms64.00110,40230.0 s
Sustained decode16K128Valid5,602.443.02,034 ms23.2 ms128.00168,06630.0 s
Sustained decode32K1Valid287.7290.0107 ms3.4 ms1.028,63130.0 s
Sustained decode32K2Valid508.2256.1117 ms3.9 ms2.0415,24630.0 s
Sustained decode32K4Valid779.3196.7193 ms5.1 ms4.0423,37930.0 s
Sustained decode32K8Valid1,145.8145.7256 ms6.9 ms8.0834,37030.0 s
Sustained decode32K16Valid1,624.9102.4482 ms9.8 ms16.0048,73530.0 s
Sustained decode32K32Valid2,341.173.2651 ms13.7 ms32.0070,23730.0 s
Sustained decode32K64Valid3,496.954.61,094 ms18.3 ms64.00104,90030.0 s
Sustained decode32K128Valid5,112.339.52,205 ms25.3 ms128.00153,34030.0 s
Sustained decode64K1Valid283.0286.2166 ms3.5 ms1.028,48830.0 s
Sustained decode64K2Valid495.5251.2200 ms4.0 ms1.9414,86730.0 s
Sustained decode64K4Valid759.8192.1294 ms5.2 ms3.9422,79430.0 s
Sustained decode64K8Valid1,100.2138.8337 ms7.2 ms7.9833,00330.0 s
Sustained decode64K16Valid1,545.996.7895 ms10.3 ms16.0046,37030.0 s
Sustained decode64K32Valid2,175.967.61,186 ms14.8 ms32.0065,27930.0 s
Sustained decode64K64Valid3,142.548.42,163 ms20.6 ms64.0094,27430.0 s
Sustained decode64K128Valid4,411.833.74,118 ms29.6 ms128.00132,39830.0 s
Sustained decode128K1Valid269.9274.9263 ms3.6 ms1.028,09930.0 s
Sustained decode128K2Valid477.9242.4334 ms4.1 ms2.0414,25629.8 s
Sustained decode128K4Valid722.3182.2341 ms5.5 ms4.0421,67030.0 s
Sustained decode128K8Valid1,013.0128.2350 ms7.8 ms7.9830,38830.0 s
Sustained decode128K16Valid1,393.086.81,333 ms11.5 ms16.0041,79330.0 s
Sustained decode128K32Valid1,881.858.12,549 ms17.2 ms32.0056,44830.0 s
Sustained decode128K64Valid2,584.839.65,114 ms25.2 ms64.0077,50730.0 s
Sustained decode128K128Valid3,413.825.89,586 ms38.8 ms128.00102,41030.0 s
Sustained decode256K1Valid248.3258.0561 ms3.9 ms0.927,44830.0 s
Sustained decode256K2Valid444.0227.6705 ms4.4 ms1.9213,31930.0 s
Sustained decode256K4Valid633.9162.3874 ms6.2 ms3.9419,01730.0 s
Sustained decode256K8Valid911.5112.91,683 ms8.9 ms8.0027,34430.0 s
Sustained decode256K16Valid1,174.772.12,931 ms13.9 ms16.0035,23230.0 s
Sustained decode256K32Valid1,517.645.85,318 ms21.8 ms32.0045,50530.0 s
Sustained decode256K64Valid1,935.028.710,372 ms34.8 ms64.0058,04530.0 s
Sustained decode256K128Valid2,370.117.221,102 ms58.0 ms128.0071,04030.0 s
Sustained decode512K1Valid219.3228.11,037 ms4.4 ms1.016,57930.0 s
Sustained decode512K2Valid385.2201.41,277 ms5.0 ms1.9211,55730.0 s
Sustained decode512K4Valid530.4139.31,242 ms7.2 ms3.9415,91130.0 s
Sustained decode512K8Valid742.092.73,569 ms10.8 ms8.0022,25630.0 s
Sustained decode512K16Valid914.756.76,628 ms17.6 ms16.0027,44230.0 s
Sustained decode512K32Valid1,093.933.412,094 ms29.9 ms32.0032,80330.0 s
Sustained decode512K64Valid1,285.619.523,207 ms51.2 ms64.0038,52830.0 s
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid170.8187.12,386 ms5.3 ms0.915,12330.0 s
Sustained decode1019K2Valid303.0167.02,613 ms6.0 ms1.929,09130.0 s
Sustained decode1019K4Valid426.5108.74,818 ms9.2 ms4.0112,79530.0 s
Sustained decode1019K8Valid529.266.68,023 ms15.0 ms8.0015,87230.0 s
Sustained decode1019K16Valid615.438.614,312 ms25.9 ms16.0018,44930.0 s
Sustained decode1019K32Valid695.821.126,748 ms47.5 ms32.0020,86530.0 s
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

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