MulaiPapan peringkatDecode calculatorModelOrganizationsReportsMatrix runsHardwareTolok UkurMarketplaceSewaProDokumen API
Bahasa
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 128 --max-num-batched-tokens 32768 --async-scheduling --mamba-backend flashinfer --mamba-ssm-cache-dtype float16 --enable-mamba-cache-stochastic-rounding --mamba-cache-philox-rounds 5 --mamba-cache-mode align --moe-backend humming --speculative_config.method dflash --speculative_config.model nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash --speculative_config.num_speculative_tokens 3

Prefill

prompt tok/s by context length
8K context37.6K218 ms TTFT
16K context41.6K390 ms TTFT
32K context43.6K739 ms TTFT
64K context41.8K1,540 ms TTFT
128K context34.8K3,693 ms TTFT
256K context25.7K9,985 ms TTFT
512K context16.7K30,721 ms TTFT
1019K context9,924103,034 ms TTFT
1024K context9,822104,621 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K293.6124 ms TTFT473.9315 ms TTFT739.6158 ms TTFT1,060.5161 ms TTFT1,629.61,221 ms TTFT2,501.31,529 ms TTFT3,855.22,709 ms TTFT5,854.14,908 ms TTFT
16K339.8122 ms TTFT511.5154 ms TTFT758.9157 ms TTFT1,009.7201 ms TTFT1,649.5798 ms TTFT2,609.11,429 ms TTFT3,782.92,593 ms TTFT5,589.54,357 ms TTFT
32K300.8204 ms TTFT478.0206 ms TTFT725.6261 ms TTFT1,019.4284 ms TTFT1,630.91,169 ms TTFT2,532.01,847 ms TTFT3,541.33,231 ms TTFT5,146.35,938 ms TTFT
64K350.8245 ms TTFT469.6249 ms TTFT726.3293 ms TTFT986.8310 ms TTFT1,502.81,251 ms TTFT2,276.42,002 ms TTFT3,185.03,393 ms TTFT4,371.96,678 ms TTFT
128K290.1375 ms TTFT524.4409 ms TTFT709.5437 ms TTFT939.3760 ms TTFT1,357.61,811 ms TTFT1,948.93,204 ms TTFT2,596.56,192 ms TTFTOver capacity
256K245.0632 ms TTFT421.5710 ms TTFT643.6681 ms TTFT843.81,783 ms TTFT1,130.33,374 ms TTFT1,518.06,409 ms TTFTOver capacityOver capacity
512K220.51,268 ms TTFT359.71,530 ms TTFT491.01,577 ms TTFT655.34,327 ms TTFT870.77,651 ms TTFTOver capacityOver capacityOver capacity
1019K207.12,687 ms TTFT278.8Requests queued2,971 ms TTFT433.24,272 ms TTFT491.08,558 ms TTFTOver capacityOver capacityOver capacityOver capacity

Requests queued: the server could not run all requests at once, so some waited in a queue. 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
Prefill8K1Valid37.6K—218 ms——1——
Prefill16K1Valid41.6K—390 ms——1——
Prefill32K1Valid43.6K—739 ms——1——
Prefill64K1Valid41.8K—1,540 ms——1——
Prefill128K1Valid34.8K—3,693 ms——1——
Prefill256K1Valid25.7K—9,985 ms——1——
Prefill512K1Valid16.7K—30,721 ms——1——
Prefill1019K1Valid9,924.0—103,034 ms——1——
Prefill1024K1Valid9,822.0—104,621 ms——1——
Sustained decode8K1Valid293.6296.1124 ms3.4 ms1.028,80730.0 s
Sustained decode8K2Valid473.9242.5315 ms4.1 ms2.0314,21530.0 s
Sustained decode8K4Valid739.6184.4158 ms5.4 ms4.0422,18530.0 s
Sustained decode8K8Valid1,060.5136.9161 ms7.3 ms8.0831,81130.0 s
Sustained decode8K16Valid1,629.6101.61,221 ms9.8 ms16.0048,89030.0 s
Sustained decode8K32Valid2,501.377.31,529 ms12.9 ms32.0174,96330.0 s
Sustained decode8K64Valid3,855.258.92,709 ms17.0 ms64.00115,64330.0 s
Sustained decode8K128Valid5,854.142.44,908 ms23.6 ms128.00175,55030.0 s
Sustained decode16K1Valid339.8325.2122 ms3.1 ms1.0210,19430.0 s
Sustained decode16K2Valid511.5251.0154 ms4.0 ms2.0415,35230.0 s
Sustained decode16K4Valid758.9185.2157 ms5.4 ms4.0422,76830.0 s
Sustained decode16K8Valid1,009.7132.5201 ms7.5 ms8.0730,28130.0 s
Sustained decode16K16Valid1,649.5102.2798 ms9.8 ms16.0149,46330.0 s
Sustained decode16K32Valid2,609.177.21,429 ms12.9 ms32.0278,23430.0 s
Sustained decode16K64Valid3,782.957.72,593 ms17.3 ms64.00113,45630.0 s
Sustained decode16K128Valid5,589.541.04,357 ms24.4 ms128.00167,58630.0 s
Sustained decode32K1Valid300.8310.4204 ms3.2 ms1.029,02530.0 s
Sustained decode32K2Valid478.0245.6206 ms4.1 ms2.0414,33830.0 s
Sustained decode32K4Valid725.6186.5261 ms5.4 ms4.0421,76630.0 s
Sustained decode32K8Valid1,019.4130.8284 ms7.6 ms8.0830,57130.0 s
Sustained decode32K16Valid1,630.998.91,169 ms10.1 ms16.0248,91530.0 s
Sustained decode32K32Valid2,532.075.21,847 ms13.3 ms32.0075,91030.0 s
Sustained decode32K64Valid3,541.353.83,231 ms18.6 ms64.00106,27430.0 s
Sustained decode32K128Valid5,146.336.45,938 ms27.5 ms128.00154,23730.0 s
Sustained decode64K1Valid350.8321.9245 ms3.1 ms1.0210,52330.0 s
Sustained decode64K2Valid469.6240.4249 ms4.2 ms2.0314,08930.0 s
Sustained decode64K4Valid726.3184.5293 ms5.4 ms4.0421,78430.0 s
Sustained decode64K8Valid986.8128.3310 ms7.8 ms8.0729,59630.0 s
Sustained decode64K16Valid1,502.894.51,251 ms10.6 ms16.0045,07130.0 s
Sustained decode64K32Valid2,276.468.82,002 ms14.5 ms32.0068,27530.0 s
Sustained decode64K64Valid3,185.047.93,393 ms20.9 ms64.0095,54730.0 s
Sustained decode64K128Valid4,371.931.36,678 ms32.0 ms128.00131,04030.0 s
Sustained decode128K1Valid290.1314.1375 ms3.2 ms1.028,70230.0 s
Sustained decode128K2Valid524.4256.0409 ms3.9 ms2.0415,73030.0 s
Sustained decode128K4Valid709.5173.3437 ms5.8 ms4.0421,28630.0 s
Sustained decode128K8Valid939.3123.2760 ms8.1 ms8.0328,17130.0 s
Sustained decode128K16Valid1,357.684.61,811 ms11.8 ms16.0040,72730.0 s
Sustained decode128K32Valid1,948.957.93,204 ms17.3 ms32.0058,45430.0 s
Sustained decode128K64Valid2,596.537.66,192 ms26.6 ms64.0077,89230.0 s
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid245.0259.0632 ms3.9 ms1.027,35030.0 s
Sustained decode256K2Valid421.5214.4710 ms4.7 ms2.0212,64230.0 s
Sustained decode256K4Valid643.6160.7681 ms6.2 ms4.0419,30930.0 s
Sustained decode256K8Valid843.8102.01,783 ms9.8 ms8.0125,30630.0 s
Sustained decode256K16Valid1,130.370.63,374 ms14.2 ms16.0033,88830.0 s
Sustained decode256K32Valid1,518.046.06,409 ms21.7 ms32.0045,53930.0 s
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid220.5233.51,268 ms4.3 ms1.016,61430.0 s
Sustained decode512K2Valid359.7191.31,530 ms5.2 ms1.9210,78930.0 s
Sustained decode512K4Valid491.0131.51,577 ms7.6 ms3.9414,72930.0 s
Sustained decode512K8Valid655.384.34,327 ms11.9 ms8.0019,65130.0 s
Sustained decode512K16Valid870.752.37,651 ms19.1 ms16.0026,09530.0 s
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid207.1226.32,687 ms4.6 ms0.916,21130.0 s
Sustained decode1019K2Requests queued278.8153.52,971 ms6.5 ms1.928,36230.0 s
Sustained decode1019K4Valid433.2111.64,272 ms9.0 ms3.9112,98930.0 s
Sustained decode1019K8Valid491.059.68,558 ms16.8 ms8.0014,72630.0 s
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

Requests queued: the server could not run all requests at once, so some waited in a queue. Over capacity: not run because this many requests at this context would exceed the server's KV cache.