H3C Tops in 25 Performance Items at the MLPerf^TM AI Benchmark Test

Recently, MLPerf™, an international authoritative AI benchmarking organization, announced the latest AI Training list, in which H3C servers R5500 G5, R5300 G5, and R4900 G5 topped the closed model partition of BERT-large (natural language processing), 3D U-Net (medical image processing), Mask R- CNN (target detection) and other 8 model tests, and won the first place in 25 configurations, demonstrating H3C's profound technical accumulation in the field of artificial intelligence.

The MLPerf™, initiated by David Patterson, winner of Turing prize, in collaboration with top academic institutions, is the most influential international AI performance benchmark. Its evaluation indicators are closely integrated with cutting-edge applications in the AI industry, and the test results have great value for technological applications and can provide authoritative and effective data for users to measure device performance. The evaluation attracted 21 mainstream chip and system manufacturers from all over the world to participate, generating a total of 264 performance scores, making the competition extremely fierce.

H3C Uniserver R5500 G5: Leading in 10 Performance Items for Super-large Model AI Training Scenarios

Equipped with NVIDIA HGX A100 8-GPU, H3C R5500 G5 server brings a 20 times AI computing capability extension, which can meet the computing challenges of super-large model AI training scenarios. In this test, under the same configuration of stand-alone and cluster mode, the R5500 G5 ranked first in 10 items among several model tests including BERT-large (natural language processing), DLRM (recommendation), Mask R-CNN (target detection), MiniGo (reinforcement learning) and 3D U-Net (medical image processing).

The R5500 G5 demonstrated outstanding performance in natural language processing, clinical diagnosis and treatment, automatic driving, and other application scenarios corresponding to BERT-large, 3D U-Net and Mask R-CNN models. 3D U-Net (medical image processing is trained based on KiTS 19 data set to find and segment cancer cells in kidneys. The model can identify whether each volume element scanned by CT is healthy or tumorous. The R5500 G5 can complete the training of 3D U-Net models on KiTS 19 data set within only 21.28 minutes, equivalent to training 288 CT images per second, which greatly improves the efficiency of medical diagnosis. This result is the best among the eight performance results of 3D U-Net models with the same configuration.

The Mask R-CNN model is applied to target detection and image processing scenarios in autonomous driving and industrial field, etc. The R5500 G5 server can complete the training of Mask R-CNN model on COCO data set containing 118,000 images within only 41.05 minutes, equivalent to completing the target detection of more than 680 images per second, thus greatly shortening the training time of AI models and further accelerates the development of AI applications.

H3C UniServer R5300 G5: Tops in 13 Performance Items, Fully Applicable to AI Reasoning/Training Scenarios

In this competition, H3C R5300 G5 server equipped with NVIDIA A100 PCIe 80GB, HGX A100 4-GPU and other AI accelerators, topped in 13 items among several models including BERT-large (natural language processing), 3D U-Net (medical image processing), Mask RCNN (target detection) and ResNet (image classification) under the same configuration.

The R5300 G5 server can support up to eight NVIDIA A30 or NVIDIA A10 GPUs in 4U space, suitable for large-scale reasoning scenarios. The R5300 G5 can also be equipped with four NVIDIA NVLink connected A100 GPUs or eight PCIe NVIDIA A100 GPUs for AI training scenarios. The agile architecture fully meets the demand for computing power resources in different scenarios.

H3C UniServer R4900 G5: Ranks First in Two Items, Featuring Low Power Consumption and High Reliability

The R4900 G5, as a mainstream two-socket racker server, took the first place in two AI training tasks among the BERT-large (natural language processing) and ResNet (image classification) models under the same configuration.

The R4900 G5 server can accommodate four NVIDIA A30 or NVIDIA A10 GPUs in a 2U space. Featuring low power consumption, high reliability, flexible scalability, and easy management, the R4900 G5 server is suitable for small-scale AI training and reasoning scenarios, providing users with higher energy efficiency and accelerating intelligent transformation.

This benchmark suite winning once again shows the solid strength of H3C in the field of AI computing. Facing the AI era under the guidance of its Cloud and AI Native strategy and its evolving “Digital Brain”, H3C will continue to focus on the needs of real application scenarios, keep improving the efficiency and capabilities of AI applications, and lead all industries toward advanced intelligence.

For more about the MLPerf™ Training v2.0 results, please visit:

https://mlcommons.org/en/news/mlperf-training-2q2022/