blobbotronic - stock.adobe.com
Vast Data storage packages Nvidia GPU for AI processing
Vast Data storage reference architecture combines the vendor's Universal Storage all-flash system with Nvidia DGX supercomputers. NFS multipathing over RDMA boosts AI data training.
Vast Data's universe has added some gravitational pull to help storage processing for AI processing.
The startup partnered with Nvidia to launch an AI reference architecture for its all-flash Universal Storage platform. The enhancement aims to help development teams boost AI inferencing.
Although billed as reference architecture, the Vast Data storage system is the same one sold as a standalone product. The new packaging does, however, give Vast Data access to Nvidia resellers and a first-mover position for Nvida's GPU DirectStorage, which is in beta.
The Vast Data storage is multi-petabyte storage geared for random access. Universal Storage targets hyperscalers and large enterprises that operate private scale-out clouds. The container-based system packages Vast Data's file servers with its LightSpeed NVMe-enabled commodity flash nodes. The network topology runs the NFS storage protocol on RDMA over Converged Ethernet to parallelize all network cards.
Vast Data serves all data from flash. Low-cost QLC SSDs provide bulk storage. Intel Optane storage class memory cards accelerate data access. Every Vast Data server has access to data in the flash pool.
The Vast Data storage controller is a Docker container. File and object protocols are accessed as microservices. The vendor claims its architecture ensures high availability with fast storage.
Regarding performance, Vast Data rates its system to feed AI workloads with up to 140 Gb/s and more than 1.5 million IOPS. The inclusion of Nvidia's DGX-based compute will push those performance numbers higher, said Jeff Denworth, Vast Data's vice president of products.
AI storage: the next growth sector
IT research firm IDC estimates unstructured data will fuel demand for AI-enabled storage technologies. IDC said files and objects will generate most new data by 2024.
"The fact that data must be handled in real time is driving significant changes in the underlying storage infrastructure requirements," Eric Burgener, an IDC research vice president for storage, said during a Vast Data-Nvidia webinar last week.
Burgener said about 70% of organizations plan to upgrade server, storage or data protection infrastructure in the next two years to support digital transformation efforts.
A few years ago, a product designed around QLC as the bulk tier would have been deemed too costly compared to using hard disk drives. However, improved endurance and lower NAND pricing gives Vast Data and other vendors a chance to innovate storage and networking topologies. Investors have injected $180 million in Vast Data since 2016, including a $100 million funding round last summer.
Most major storage vendors offer versions of their arrays that include Nvidia GPUs. Dell EMC added an all-flash model to its Isilon scale-out NAS lineup, while flash specialist Pure Storage partnered earlier with Nvidia on Pure AIRI, which injects DGX compute in Pure's NVMe-based FlashBlade NAS.
Nvidia DGX is part of a reference architecture put together by NetApp and Lenovo. That product incorporates NetApp's OnTap storage operating system and Lenovo ThinkSystem servers.
Vast Data said its LightSpeed enclosure will support Nvidia GDS. Nvidia has yet to specify a date for general availability. Five mutual resellers will set pricing for the Vast Data-Nvidia reference design: Cambridge Computer, Mark III Systems, Trace3, Uclick and Xenon.
In addition to storage partnerships, Nvidia made other storage-related moves last year, acquiring object storage startup SwiftStack and storage networking company Mellanox Technologies.
