Menu
Back previous/next report: k j

NVIDIA NVIDIA Vera Rubin NVL72 (4-GPU SUT)

Submitted by NVIDIAon 2026-10-17 Published on 2026-10-19

Result statement

NVIDIA ran deepseek-v4_1-flash on NVIDIA Vera Rubin NVL72 (4-GPU SUT) and achieved a provisional 103,537.0 tokens/second System TPS at 7,085.1ms TTFT P90 in the Standardized division of the Available category on October 19, 2026. Pending MLCommons Association verification, results are subject to change pending peer review. .

[1] MLPerf Endpoints v1.0 deepseek-v4_1-flash, published October 19, 2026; retrieved from endpoints.mlcommons.org on November 7, 2026. Result —, run 10: concurrency = 3,072 queries, TTFT P90 = 7,085.1ms, System TPS = 103,537.0 tokens/sec, interactivity = 33.7 tokens/sec per user and total provisioned system power = 12.61 kW. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Submission Info
NVIDIA
618254e9-a8b8-47e6-8851-d21dd6b4af51
Submission Status REVIEW_PENDING
2026-10-17
2026-10-19
Division Standardized
Deployment Client on Premises
Model
Model Name deepseek-v4_1-flash
Division Standardized
Model Precision —
Model Link —
Calibration Link —
Model Notes —
Datasets
Dataset Name mlperf_agentic_inference_dataset
Dataset Type performance
Average Input Tokens —
Dataset Link —

System TPS vs. Tokens/s per User

020k40k60k80k100kSystem TPS50100150200250Tokens/s per User

System TPS vs. Concurrency

020k40k60k80k100kSystem TPS01,0002,0003,000# of queries

Time to First Token P90 vs. Concurrency

02.0k4.0k6.0kTTFT P90 (ms)01,0002,0003,000# of queries

Tokens/s per User vs. Concurrency

50100150200250Tokens/s per User01,0002,0003,000# of queries
SUT Summary
NVIDIA
NVIDIA Vera Rubin NVL72 (4-GPU SUT)
Available
System Category — · —
4 accelerators
1
1
Max Supported Concurrency 3,072
Node Type ID 0
Node Count 1
Software
Framework vLLM 0.20.2rc1.dev6106+g7cf5542c2 behind Dynamo 1.6.0.dev20260924
vLLM 0.20.2rc1.dev6106+g7cf5542c2 behind Dynamo 1.6.0.dev20260924
vLLM behind a Dynamo frontend
620
Link to Container registry.gitlab.com/nvidia/mlperf-inference-partner/nv-mlpinf-partner/centml-rubin-vllm:dsv41-rubin-7cf5542c25-r1
Operating System unspecified
Filesystem unspecified
Other Software unspecified
Software Notes Each vLLM worker runs on 12 dedicated physical cores of its GPU's CPU socket with its memory bound there (numactl --physcpubind/--membind); the Dynamo frontend and the endpoint client run on the remaining CPUs. Dynamo frontend tokenizer prefix cache 16 GB (C2048) / 32 GB (C3072) with DYN_TOKENIZER_CACHE_EXTEND=1. Rubin kernels: FlashMLA sm_107a, DeepGEMM MegaMoE SM107 with per-die routing (branch dsv41-rubin-megamoe-sm107), Rubin MXFP8 dense GEMM dispatch.
Hardware

Processor

NVIDIA Vera CPU
2
176
VCPUs Per Processor 352

Accelerator

—
0
—
—
—
—

Host / Storage

unspecified
Memory Configuration unspecified
Storage Capacity unspecified
Storage Type unspecified

Networking

unspecified
NVLink + InfiniBand
Network Topology —

Other

Other Hardware —
Cooling Liquid-cooled
Hardware Notes —
Run Data
Field NameRun 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8Run 9Run 10
Run Date10/17/202610/17/202610/17/202610/17/202610/17/202610/17/202610/17/202610/17/202610/17/202610/17/2026
Concurrency162432641282565121,0242,0483,072
System Tokens/Second3,741.65,266.26,643.812,000.320,302.733,555.153,781.875,011.793,065.5103,537.0
Tokens/Second per User233.9219.4207.6187.5158.6131.1105.073.345.433.7
TTFT P90 (ms)112.8114.9118.8100.9145.2196.6327.72561.72424.87085.1
Utilization3.6%5.1%6.4%11.6%19.6%32.4%51.9%72.4%89.9%100.0%
Configuration Summarydeepseek-v4.1-flash-nvflow · online · c16deepseek-v4.1-flash-nvflow · online · c24deepseek-v4.1-flash-nvflow · online · c32deepseek-v4.1-flash-nvflow · online · c64deepseek-v4.1-flash-nvflow · online · c128deepseek-v4.1-flash-nvflow · online · c256deepseek-v4.1-flash-nvflow · online · c512deepseek-v4.1-flash-nvflow · online · c1024deepseek-v4.1-flash-nvflow · online · c2048deepseek-v4.1-flash-nvflow · online · c3072
Config Typeonlineonlineonlineonlineonlineonlineonlineonlineonlineonline
Config Region——————————
Target QPS——————————
Result Statecompletecompletecompletecompletecompletecompletecompletecompletecompletecomplete
Measured Accuracy Score——————————
Benchmark Versionunknownunknownunknownunknownunknownunknownunknownunknownunknownunknown
Result Version0.1.00.1.00.1.00.1.00.1.00.1.00.1.00.1.00.1.00.1.0
Archive URIgs://mlperf-endpoints-blob/runs/f8a89260-f03d-4299-8b13-1f777af24d6c/archive.tar.gzgs://mlperf-endpoints-blob/runs/6bb91092-fc26-4848-aaf0-157461dbe715/archive.tar.gzgs://mlperf-endpoints-blob/runs/4cd88602-aed8-4aa3-b21c-81d280d13f37/archive.tar.gzgs://mlperf-endpoints-blob/runs/6c0a4110-464c-4310-8b9d-3e48b80f68e8/archive.tar.gzgs://mlperf-endpoints-blob/runs/860a87b6-20d8-4f59-9209-0b49f04998a5/archive.tar.gzgs://mlperf-endpoints-blob/runs/fe4b1e32-04c3-4213-b58e-ca8bedd2f999/archive.tar.gzgs://mlperf-endpoints-blob/runs/d0b5df69-5a34-4fd4-b0a1-49e297fc604f/archive.tar.gzgs://mlperf-endpoints-blob/runs/75bb614f-c593-4488-8c53-c3a8de63a789/archive.tar.gzgs://mlperf-endpoints-blob/runs/1a5e5737-f9c8-46b4-a1c8-e906500e8217/archive.tar.gzgs://mlperf-endpoints-blob/runs/285def9f-29a8-4c67-91b2-c6a8b02add04/archive.tar.gz
QPS4.07045.80077.386512.944822.344136.516758.504080.685598.8827107.5036
Total Output Tokens18,030,85617,507,24517,109,48836,297,04671,524,422105,208,158136,691,236196,761,109246,104,851360,235,195
Average Output Tokens919.24907.86899.46927.03908.64918.9919.28929.68941.17963.1
Run Duration (s)4,818.983,324.432,575.253,024.693,522.903,135.382,541.592,623.072,644.433,479.29
Total Queries19,61519,28419,02239,15478,716114,494148,693211,644261,488374,036
Total Queries Issued19,61519,28419,02239,15478,716114,494148,693211,644261,488374,036
Total Queries Failed0000000000
Time to First Token (TTFT) (ms)
Minimum28.228.528.930.633.743.949.981.6114.5778.4
Average80.371.896.370.198.7150.0301.51316.01584.75459.5
P5057.858.759.760.079.8100.0139.1871.11075.54847.8
P90112.8114.9118.8100.9145.2196.6327.72561.72424.87085.1
P95150.0148.4154.3124.6174.5241.8439.53094.82890.17776.0
P99240.7230.8284.1188.5293.2555.31068.34725.219150.027747.7
P9994243.3940.510984.21166.23473.810627.727670.343363.840455.156189.3
Maximum13483.21920.919071.22449.74707.513168.336462.848733.448501.964178.6
Time Per Output Token (TPOT) (ms)
Minimum0.830.900.960.931.380.070.010.010.01-0.09
Average2.012.172.382.853.594.545.858.0613.115.8
P501.832.012.152.643.334.175.347.4512.115.1
P902.863.153.414.165.246.738.6612.021.124.5
P953.123.433.734.525.697.349.4213.023.427.2
P993.754.084.545.386.748.6211.015.328.032.9
P99917.46.3127.712.011.517.527.325.539.249.4
Maximum70.935.389.945.643.473.8186.4192.22125.7493.1
Query Latency (ms)
Minimum63.268.875.473.890.594.4158.2201.7390.81632.8
Average2314.12489.42681.33315.04106.05316.66953.210632.717803.724979.8
P50413.9443.3480.1555.5710.4902.11206.52867.53752.18352.1
P905411.65740.76226.87889.19529.012494.116895.224949.641393.255179.4
P9516502.716951.417801.823487.627370.136402.349132.571902.0132437.9164465.1
P9924345.526798.328910.835268.944406.456978.273428.0102736.3185125.7219413.4
P99926812.628868.131742.037741.547276.560853.079011.7110122.8202695.6240529.5
Maximum58181.632531.159813.946012.453482.674823.9115068.1142374.8225866.8276666.1