Per iniziareClassificaDecode calculatorModelliOrganizationsReportsMatrix runsHardwareValutazioni comparativeMarketplaceNoleggiProDocumentazione API
Lingua
Actual Computer — Every computer, one endpoint

Speed test matrix

NVIDIA-Nemotron-3-Super-120B-A12B-FP8 on H200 NVL 141GB ×2

vllm · 0.26.1rc1.dev638+g52be12cfa · 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
0.26.1rc1.dev638+g52be12cfa
Quantization
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

env VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 "$HOME/env_h200-nvidia-nemotron/bin/vllm" serve nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8 --served-model-name nemotron --trust-remote-code --tensor-parallel-size 2 --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 1048576 --gpu-memory-utilization 0.880548 --host 0.0.0.0 --port 8000 --api-key <redacted> --async-scheduling --dtype auto --kv-cache-dtype fp8 --max-cudagraph-capture-size 128 --enable-chunked-prefill --mamba-ssm-cache-dtype float32

Prefill

prompt tok/s by context length
8K context30.3K270 ms TTFT
16K context30.5K531 ms TTFT
32K context31.5K1,024 ms TTFT
64K context29.6K2,177 ms TTFT
128K context27K4,760 ms TTFT
256K context22.8K11,260 ms TTFT
512K context17.2K29,899 ms TTFT
1019K context11.5K88,645 ms TTFT
1024K context11.5K89,383 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K140.9252 ms TTFT267.2379 ms TTFT475.9596 ms TTFT767.71,041 ms TTFT1,170.51,938 ms TTFT1,775.43,972 ms TTFT2,854.07,568 ms TTFT4,357.414,899 ms TTFT
16K140.3274 ms TTFT266.9400 ms TTFT478.7638 ms TTFT764.51,110 ms TTFT1,155.42,048 ms TTFT1,786.53,887 ms TTFT2,779.87,356 ms TTFT4,222.514,607 ms TTFT
32K139.5311 ms TTFT265.3424 ms TTFT472.5693 ms TTFT761.51,197 ms TTFT1,151.12,195 ms TTFT1,761.03,922 ms TTFT2,726.17,402 ms TTFT4,063.614,668 ms TTFT
64K137.6360 ms TTFT261.5463 ms TTFT469.9798 ms TTFT767.41,360 ms TTFT1,123.62,235 ms TTFT1,698.53,847 ms TTFT2,583.47,153 ms TTFT3,757.614,001 ms TTFT
128K134.2427 ms TTFT255.7508 ms TTFT462.2741 ms TTFT735.51,176 ms TTFT1,069.81,904 ms TTFT1,586.03,230 ms TTFT2,336.55,766 ms TTFT3,239.411,120 ms TTFT
256K126.2897 ms TTFT249.51,121 ms TTFT446.91,586 ms TTFT682.32,532 ms TTFT989.34,126 ms TTFT1,395.97,813 ms TTFT1,925.514,544 ms TTFTOver capacity
512K118.91,650 ms TTFT229.52,050 ms TTFT410.12,945 ms TTFT600.34,686 ms TTFT854.78,129 ms TTFT1,116.015,084 ms TTFTOver capacityOver capacity
1019K102.03,585 ms TTFT197.44,136 ms TTFT358.15,668 ms TTFT498.38,991 ms TTFT655.015,533 ms TTFTOver 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
Prefill8K1Valid30.3K—270 ms——1——
Prefill16K1Valid30.5K—531 ms——1——
Prefill32K1Valid31.5K—1,024 ms——1——
Prefill64K1Valid29.6K—2,177 ms——1——
Prefill128K1Valid27K—4,760 ms——1——
Prefill256K1Valid22.8K—11,260 ms——1——
Prefill512K1Valid17.2K—29,899 ms——1——
Prefill1019K1Valid11.5K—88,645 ms——1——
Prefill1024K1Valid11.5K—89,383 ms——1——
Sustained decode8K1Valid140.9142.3252 ms7.0 ms1.014,22830.0 s
Sustained decode8K2Valid267.2134.9379 ms7.4 ms1.928,01430.0 s
Sustained decode8K4Valid475.9118.2596 ms8.5 ms4.0014,27730.0 s
Sustained decode8K8Valid767.794.31,041 ms10.6 ms8.0023,02330.0 s
Sustained decode8K16Valid1,170.570.41,938 ms14.2 ms16.0035,11030.0 s
Sustained decode8K32Valid1,775.451.93,972 ms19.3 ms32.0053,25030.0 s
Sustained decode8K64Valid2,854.037.97,568 ms26.4 ms64.0085,63330.0 s
Sustained decode8K128Valid4,357.425.114,899 ms39.8 ms128.00130,82530.0 s
Sustained decode16K1Valid140.3141.9274 ms7.0 ms1.014,21030.0 s
Sustained decode16K2Valid266.9135.4400 ms7.4 ms2.028,00730.0 s
Sustained decode16K4Valid478.7118.8638 ms8.4 ms4.0014,35930.0 s
Sustained decode16K8Valid764.594.11,110 ms10.6 ms8.0022,92930.0 s
Sustained decode16K16Valid1,155.469.82,048 ms14.3 ms16.0034,65730.0 s
Sustained decode16K32Valid1,786.551.63,887 ms19.4 ms32.0053,59930.0 s
Sustained decode16K64Valid2,779.837.27,356 ms26.9 ms64.0083,39830.0 s
Sustained decode16K128Valid4,222.524.414,607 ms40.9 ms128.00126,63530.0 s
Sustained decode32K1Valid139.5141.4311 ms7.1 ms1.014,18330.0 s
Sustained decode32K2Valid265.3134.7424 ms7.4 ms1.927,96430.0 s
Sustained decode32K4Valid472.5117.5693 ms8.5 ms4.0014,18030.0 s
Sustained decode32K8Valid761.593.91,197 ms10.7 ms8.0022,84530.0 s
Sustained decode32K16Valid1,151.169.72,195 ms14.3 ms16.0034,52830.0 s
Sustained decode32K32Valid1,761.051.03,922 ms19.6 ms32.0052,83230.0 s
Sustained decode32K64Valid2,726.136.27,402 ms27.7 ms64.0081,78830.0 s
Sustained decode32K128Valid4,063.623.614,668 ms42.5 ms128.00122,10730.0 s
Sustained decode64K1Valid137.6140.0360 ms7.1 ms1.014,12730.0 s
Sustained decode64K2Valid261.5132.8463 ms7.5 ms2.027,84230.0 s
Sustained decode64K4Valid469.9117.0798 ms8.5 ms4.0014,09630.0 s
Sustained decode64K8Valid767.494.91,360 ms10.5 ms8.0023,01530.0 s
Sustained decode64K16Valid1,123.668.32,235 ms14.6 ms16.0033,69530.0 s
Sustained decode64K32Valid1,698.549.63,847 ms20.2 ms32.0050,94330.0 s
Sustained decode64K64Valid2,583.434.67,153 ms28.9 ms64.0077,50030.0 s
Sustained decode64K128Valid3,757.621.914,001 ms45.6 ms128.00112,77230.0 s
Sustained decode128K1Valid134.2137.7427 ms7.3 ms1.014,02530.0 s
Sustained decode128K2Valid255.7130.0508 ms7.7 ms1.927,67030.0 s
Sustained decode128K4Valid462.2115.0741 ms8.7 ms4.0013,86330.0 s
Sustained decode128K8Valid735.591.01,176 ms11.0 ms8.0022,06130.0 s
Sustained decode128K16Valid1,069.865.51,904 ms15.3 ms16.0032,09630.0 s
Sustained decode128K32Valid1,586.047.03,230 ms21.3 ms32.0047,58130.0 s
Sustained decode128K64Valid2,336.532.55,766 ms30.8 ms64.0070,08030.0 s
Sustained decode128K128Valid3,239.420.211,120 ms49.4 ms128.0097,15330.0 s
Sustained decode256K1Valid126.2135.1897 ms7.4 ms1.013,78730.0 s
Sustained decode256K2Valid249.5124.31,121 ms8.0 ms1.927,31629.3 s
Sustained decode256K4Valid446.9110.01,586 ms9.1 ms4.0013,40330.0 s
Sustained decode256K8Valid682.382.32,532 ms12.1 ms8.0020,46330.0 s
Sustained decode256K16Valid989.357.34,126 ms17.5 ms16.0029,67730.0 s
Sustained decode256K32Valid1,395.937.57,813 ms26.7 ms32.0041,85530.0 s
Sustained decode256K64Valid1,925.522.514,544 ms44.4 ms64.0057,76730.0 s
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid118.9119.61,650 ms8.4 ms1.003,56730.0 s
Sustained decode512K2Valid229.5114.42,050 ms8.7 ms2.006,88630.0 s
Sustained decode512K4Valid410.1100.62,945 ms9.9 ms4.0012,30430.0 s
Sustained decode512K8Valid600.371.74,686 ms13.9 ms8.0018,00730.0 s
Sustained decode512K16Valid854.748.48,129 ms20.7 ms16.0025,64030.0 s
Sustained decode512K32Valid1,116.028.915,084 ms34.6 ms32.0033,46930.0 s
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid102.0103.03,585 ms9.7 ms1.003,06030.0 s
Sustained decode1019K2Valid197.497.94,136 ms10.2 ms2.005,92430.0 s
Sustained decode1019K4Valid358.186.05,668 ms11.6 ms4.0010,74030.0 s
Sustained decode1019K8Valid498.356.88,991 ms17.6 ms8.0014,94430.0 s
Sustained decode1019K16Valid655.034.115,533 ms29.3 ms16.0019,64930.0 s
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.