H3C Tops in 61 Performance Items at the MLPerfTM AI Benchmark Test23-09-2022
H3C servers came out first in 61 items for Datacenter and Edge scenario tasks in the latest MLPerf™ AI Inference v2.1 list, which received a total of 5,300 performance results submitted by 21 mainstream chip, system, and cloud platform manufacturers.
In June, H3C servers topped the MLPerf™ lists of 25 performance items for AI training scenarios. The latest remarkable performance results have once again demonstrated H3C's extraordinary strength in the field of artificial intelligence.
The MLPerf™, initiated by David Patterson, winner of Turing prize, in collaboration with top academic institutions including Stanford University and Harvard University, is the most influential international AI performance benchmark. Its evaluation indicators are closely integrated with cutting-edge applications in the AI industry, and cover major AI application scenarios, such as BERT (natural language processing), DLRM (recommendation), ResNet (image classification), and 3D U-Net (medical image processing), which helps users better understand manufacturers' strengths in the AI industry and provides useful guidelines that facilitate the development and application of AI technologies.
Datacenter Scenario Inference Tasks: 34 World Champions
The latest round of MLPerf Inference benchmark test included 16 scenario-based test items under six models designed for various computing scenarios, including 3D U-Net (medical image processing), RetinaNet (object detection), and ResNet50 (image classification). These systems are mainly used in such scenarios as medical image processing, automaticdriving, and industrial quality testing, which are typical fields where data centers are used and represent the main battlefield with the fiercest competition in the MLPerf™ tests. H3C's AI servers ranked first in 34 items in the MLPerf™ Inference benchmark test for application scenarios of data centers, including one best result in terms of configuration.
In the task under the 3D U-Net model, H3C UniServer R5300 G5 handled the processing of 13.04 3D medical images per second with 99.9% accuracy. It can help doctors complete analysis of nidus of disease within less than 20 seconds.
In the task under the ResNet50 model, H3C R5500 G5 classified 314,368 images per second.
In the task under the RetinaNet model, H3C R5500 G5 finished target detection in 4,657.04 images per second.
As an all-rounder for AI inference and training scenarios, H3C UniServer R5300 G5 recorded a total of 25 best results among those of servers with the same configuration, demonstrating its capacity for supporting large-scale, diversified, and highly complex AI scenarios.
The server can provide multiple types of topology configuration, including 1:8, 1:4, and 1:2, to meet the needs of different AI scenarios. Besides, a single H3C UniServer R5300 G5 can support up to eight dual-width GPUs or 20 single-width GPUs, which enables it to reach an operating speed of 40 traditional servers combined.
Edge Scene Inference Track: 27 World Champions
H3C's three AI servers- H3C UniServer R5500 G5, R5300 G5, and R4900 G5- generated 27 best results in tasks for edge scenarios, including 3 absolute configuration champions. H3C UniServer R5500 G5, in particular, outperformed other manufacturers’ servers with the same configuration to record 19 best results, demonstrating great competitive strengths in edge inference scenarios.
H3C UniServer R5500 G5 has optimized the Transformer model and is able to provide strong support for many AI services, including machine vision and natural language processing. It can guarantee BERT inference with a time-lag of merely 1.53 ms in single-threaded process. In addition, the R5500 G5 server can also support the new Multi-Instance GPU feature, which allows a GPU to be partitioned into seven separate GPU instances, with each running different applications. The feature has significantly improved GPU's utilization rate and can meet users' need for fast business deployment.
Facing the AI era, H3C will continue to focus on the needs of real application scenarios under the guidance of its Cloud and AI Native strategy and its evolving “Digital Brain” initiative, keep improving the efficiency and capabilities of AI applications, and lead all industries toward advanced intelligence and inject impetus into the development of the digital economy.