Menu
Back previous/next report: k j

NVIDIA NVIDIA H200 (8x H200-SXM-141GB, TensorRT)

Submitted by NVIDIA

Submission Info
Submitter NVIDIA
Submission ID nvidia-H200-SXM-141GBx8_TRT
Submission Status
Submission Date
Publication Date
Division Closed
Model
Model Name Deepseek-R1
Division Closed
Model Precision FP4
Model Link
Calibration Link
Model Notes
Datasets
Dataset Name MLPerf-Deepseek
Dataset Type Performance
Average Input Tokens 803.791249
Dataset Link

System TPS vs. Tokens/s per User

0204060System TPS20406080100Tokens/s per User

System TPS vs. Concurrency

0204060System TPS024# of queries

Time to First Token P95 vs. Concurrency

050100150200TTFT P95 (s)024# of queries

Tokens/s per User vs. Concurrency

20406080100Tokens/s per User024# of queries
SUT Summary
NVIDIA
System Name NVIDIA H200 (8x H200-SXM-141GB, TensorRT)
System Availability Available
System Category Datacenter · On-premise
System Size 8x H200
Node Types 1
Total Node Count 1
Max Supported Concurrency
Node Type ID 1
Node Count 1
Software
Ubuntu
Framework TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0
Serving Framework TensorRT 10.14, CUDA 13.1, cuDNN 9.17, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0
Inference Backend
Driver
Link to Container
Operating System Ubuntu 24.04
Filesystem
Other Software TensorRT 10.14, CUDA 13.1, cuDNN 9.17, Driver 580.105, TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0
Software Notes
Hardware

Processor

Processor Model Name Intel(R) Xeon(R) Platinum 8480C
Processors per Node 2
Cores Per Processor 56
VCPUs Per Processor

Accelerator

Accelerator Model Name NVIDIA H200-SXM-141GB
Accelerators per Node 8
Memory Type HBM3e
Memory Capacity 141 GB
Accelerator Interconnect 18x 4th Gen NVLink, 900GB/s
Host-Accelerator Interconnect PCIe Gen5 x16

Host / Storage

Host Memory Capacity 2 TB
Memory Configuration 32x 64GB MTC40F2046S1RC48BA1
Storage Capacity 2 TB SSD, 5 TB CIFS
Storage Type NVMe SSD, CIFS mounted disk storage

Networking

Network Interfaces 10x 400Gbe Infiniband
Network Details Infiniband; Data bandwidth for GPU-PCIe: 504GB/s; PCIe-NIC: 500GB/s
Network Topology

Other

Other Hardware
Cooling Air-cooled
Hardware Notes H200 TGP 700W
Run Data
Field NameRun 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8Run 9Run 10
Run Date03/03/202603/02/202603/03/202603/03/202603/03/202602/27/202602/27/202602/27/202602/27/202602/27/2026
Concurrency0000001135
System Tokens/Second1.01.72.94.77.29.816.027.445.268.2
Tokens/Second per User93.081.168.756.543.029.524.120.517.012.8
TTFT P95 (ms)11652.213152.713033.613414.923136.422800.728881.043784.280819.4227214.5
Utilization1.4%2.5%4.2%6.9%10.5%14.4%23.5%40.2%66.3%100.0%
Configuration SummaryTP=8, PP=1, batch=64, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=64, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=64, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=64, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=64, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=128, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=128, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=128, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=128, precision=FP8, kv_cache=fp8, cont_batchTP=8, PP=1, batch=128, precision=FP8, kv_cache=fp8, cont_batch
Config Type
Config Region
Target QPS
Result State
Measured Accuracy Score
Benchmark Version
Result Version
Archive URI
QPS0.00020.00040.00060.00120.00180.00260.00430.00740.01190.0176
Total Output Tokens168,384315,398539,431867,0741,378,2141,986,4753,400,8936,215,38410,495,59916,968,991
Average Output Tokens
Run Duration (s)1,810.331,944.101,962.851,919.712,003.162,102.162,209.292,363.312,418.592,594.20
Total Queries29681172203395169061,6752,7594,388
Total Queries Issued
Total Queries Failed
Time to First Token (TTFT) P95 (ms)
Minimum10084.611549.311442.711796.312352.212501.713088.813671.015126.018639.0
Average10795.712495.612562.013372.013652.414241.018518.620908.927876.853407.6
P5010720.612255.512240.212618.512996.313499.914232.415370.917820.725893.4
P9011219.512931.712861.413184.813533.814807.620993.925002.679431.5204936.4
P9511652.213152.713033.613414.923136.422800.728881.043784.280819.4227214.5
P9912239.420479.520962.022399.923686.122979.229220.956258.185391.5229120.7
P99912371.020609.120968.090508.423709.523765.6691087.3564477.9550733.9487221.1
Maximum12385.720623.520968.3109605.023711.023770.0693520.9569426.8855644.6651657.8
Time Per Output Token (TPOT) (ms)
Minimum988.81091.91203.71379.81699.22315.82786.72794.63420.84031.7
Average1010.01134.71330.71646.62111.22941.33494.03916.04663.55905.0
P50998.91140.41328.61648.72121.92952.33440.93835.94557.85834.7
P901043.01163.11362.31690.12155.92970.73509.14044.54953.86293.7
P951049.81170.41367.21694.92164.52976.63891.34445.95318.76566.8
P991052.81175.61387.21707.02171.52993.24926.45977.36502.47241.6
P9991053.21179.01396.61709.62176.83003.76862.77639.97895.48314.0
Maximum1053.31179.41397.61710.02178.83010.19007.59847.110867.78743.1
Query Latency (ms)
Minimum927674.4788299.1837363.5644657.7699664.21135927.9951145.9997340.1871014.81150544.0
Average5990205.15315279.56154659.76499534.18548891.911276621.613056397.514415557.617680229.822647036.3
P502422761.52545455.13101725.43244005.14457795.46154640.87076876.17683333.89498088.012149481.9
P9015335188.215088890.314485372.615789616.222644300.326279624.531251691.236287802.443714543.656870900.0
P9518599786.818380822.723198288.123184326.531616174.840386936.244238223.051368579.763169293.079066406.2
P9920704536.521001481.227152850.433429049.542396244.858646899.472150174.577121298.291616497.0116331712.7
P99921039690.823019464.427241985.633726948.142875480.459112422.873155384.180648155.594797002.9121436002.2
Maximum21076930.223243684.727247533.433789303.942989680.959185143.473162326.480685205.196152892.2121712265.4