Huawei AI Storage Ranks First in Global Performance in 2024 MLPERF™ AI Benchmark

MLPERF, the world's most authoritative and influential AI benchmark test, released the results of the Storage V1.0 AI benchmark test. Huawei's new AI storage OceanStor A800 stood out among global manufacturers, with total performance, per-node performance, and per-U performance all ranking first in the world.

MLPerf benchmark is a standardized basic test platform for measuring the performance of artificial intelligence hardware, software and services. It was founded by Turing Award winner David Patterson in collaboration with top companies and academic institutions such as Google, Stanford, and Harvard University. It is the most authoritative and influential international AI performance benchmark. In order to keep up with the forefront of the industry, MLPERF releases benchmark data every year and has been widely recognized around the world. Among them, the 3D U-Net workload test matches the industry's trend of multimodal large models, has the highest requirements for storage bandwidth, and can more comprehensively and accurately reflect the performance of storage in large-scale AI clusters.

A total of 13 mainstream manufacturers in the industry participated in this storage performance test. By running a distributed training test program to simulate the GPU computing process, the access model of the AI ​​server to the storage system was restored to the greatest extent. The storage system test performance was evaluated by simulating the maximum number of GPUs that the AI ​​storage system can support. Among them, the 3D U-Net workload requires the highest storage bandwidth per FLOPS computing power, and the data needs to be read from the storage node, and it is not allowed to be cached on the host in advance. It can truly reflect the performance of the storage system and actually reflect the large model storage experience.

Huawei OceanStor A800, the AI ​​storage that won the world's first place in the storage performance test, successfully met the data throughput requirements of 255 GPU simulation training with a single device, with GPU utilization rate maintained at over 90%, and a single-frame stable bandwidth of up to 679 GB/s, which is 10 times the performance of traditional storage. It provides up to 340GB/s bandwidth per node and 85GB/s bandwidth per U on average, which are 2.88 times and 1.44 times the second place respectively, reflecting the comprehensive leading competitiveness of OceanStor A800. At the same time, OceanStor A800 can provide hundreds of TB-level bandwidth through large-scale horizontal expansion, and the checkpoint read and write time is shortened from 10 minutes to seconds, and the breakpoint continuation time is less than 15 minutes, which greatly reduces GPU waiting, increases end-to-end computing power utilization by 30%+, and comprehensively improves the efficiency of large model training. It topped the MLPerf Storage V1.0 benchmark performance list for the first time. This achievement not only demonstrates the outstanding strength of Huawei's data storage technology, but also marks another key milestone for Huawei's data storage in the industry.

Huawei has been working hard on data storage and continuously innovating, launching the new OceanStor A series AI storage. Aiming at mixed loads in AI scenarios, based on brand-new hardware, it has built a storage system with leading architecture, featuring excellent performance, EB-level elastic expansion, and long memory for inference, which can fully accelerate the training and promotion of large AI models.

Looking to the future, Huawei AI Storage will continue to deepen its research in the field of AI big models, constantly break through performance limits, ride the wave of AI big model data, and lead the future of data.