Huawei Reveals OceanStor A310: A Rapid Storage Solution Tailored for AI Model Trainers

Crypto News– Huawei Unveils OceanStor A310 at GITEX GLOBAL 2023, an Innovative AI Storage Solution for Enhanced Large Model Applications. The OceanStor A310 is designed to address challenges in the AI industry by offering a high-speed storage solution for a range of scenarios, from basic model training to industry model training and inference.

Huawei Reveals OceanStor A310: A Rapid Storage Solution Tailored for AI Model Trainers

Imagine the OceanStor A310 as an incredibly efficient librarian in a vast digital library, rapidly retrieving information. In comparison, IBM’s ESS 3500 acts like a slower librarian. The faster retrieval capabilities of OceanStor A310 enable AI applications to work swiftly, facilitating quick and intelligent decision-making. This remarkable speed in accessing data sets OceanStor A310 apart.

What sets the OceanStor A310 apart is its ability to accelerate data processing for AI. In a comparison with IBM’s ESS 3500, Huawei’s latest all-flash array reportedly boosts the feeding speed of Nvidia GPUs by nearly four times on a per-rack unit basis. This achievement is attributed to the methodology of using Nvidia’s Magnum GPU Direct, where data is transferred directly from NVMe storage resources to GPUs without the involvement of a storage host system.

In order to address this challenge effectively, each OceanStor unit has the capability to accommodate as many as 96 NVMe SSDs, processors, and a memory cache. Furthermore, users have the flexibility to cluster up to 4,096 A310 units, all of which share a global file system supporting standard application protocols. The OceanStor A310 is primarily designed to minimize data transmission time, leveraging SmartNICs and a massively parallel architecture.

Block and Files, in a benchmark study comparing Huawei’s solution with its direct competitors, stated, “Huawei’s A310, featuring compact nodes, outperformed all others in terms of both sequential reading and writing speeds. It achieved an impressive 41.6/80GBps sequential write/read bandwidth, surpassing IBM’s 30/63GBps figures.”