Nabugu - stock.adobe.com

Vast Data, Nvidia collaborate on new AI architecture

Vast Data is adding its OS to Nvidia hardware, freeing up GPUs to focus solely on AI workloads -- a new architecture that could potentially speed up processing.

Vast Data is introducing a new cloud AI architecture that uses its parallel file software on Nvidia hardware and is being deployed by CoreWeave. The architecture is designed to increase performance on GPU clusters while providing multi-tenancy for customers.

Ahead of next week's Nvidia GTC event, Vast Data unveiled that it has taken its parallel file operating system and placed it on the latest Nvidia BlueField-3 data processing units (DPUs). The DPUs will offload storage functions and improve security and multi-tenancy for users, leaving the GPUs to process AI workloads.

This use of DPUs is focusing on the promises of the technology, according to Steve McDowell, an analyst and founding partner at NAND Research. Vast's software typically uses a dedicated server, but in this architectural design with certain tasks being offloaded to the DPU and the DPU being able to directly communicate with GPU clusters, a separate server is not necessary.

"This keeps that machine free to do AI [workloads]," he said.

Utilizing DPUs for AI

The new architecture will be first deployed by CoreWeave, a GPU cloud service provider Vast began partnering with in September 2023. The BlueField-3 DPUs increase efficiency of the cluster by offloading data processing, which means fewer x86 servers are needed for I/O, according to Vast.

Maximum GPU performance is typically gained by giving the user root access to a physical server, according to John Mao, vice president of global business development at Vast Data. This allows everything to be seen on the back end, which isn't ideal for security for either service providers or customers. With the Vast operating system on a DPU, there is a level of insulation for both the customer and service provider, as they still have root access but only through the DPU.

Anything that makes [multi-tenancy] simpler is better for service providers.
Steve McDowellAnalyst, NAND Research

Large GPU clusters from cloud providers including CoreWeave are largely shared between multiple customers, McDowell said. Even internal GPU clusters are shared among different teams, making multi-tenancy a priority.

"Anything that makes [multi-tenancy] simpler is better for service providers," he said.

The architecture is beneficial as it physically isolates the software stack as well, McDowell said. This allows the GPU compute to get to a customer without exposing what the customer is running.

Expanding hardware partners

Along with introducing a new AI architecture, Vast is also now partnering with Supermicro on a full-stack AI offering -- from storage to compute. This aims to give server providers and hyperscalers high-performance software and hardware combinations for AI workloads that use Vast software and Supermicro hardware, and are certified by Nvidia.

Vast is a software company but has sold its product on a specially made hardware array, Ceres. The array uses a combination of BlueField-2 DPUs, storage class memory and quad-level cell SSD technology. Since then, it has begun broadening its hardware approach. In April 2023, Vast software began to ship on HPE Alletra MP hardware, Vast's first hardware partnership with a legacy storage player. With Supermicro, Vast is moving more firmly into software-defined storage territory as it relies less on specific hardware, according to Mitch Lewis, an analyst at Futurum Group.

"Software-defined storage has risen in popularity due to the flexibility it provides," he said. "However, many customers ultimately prefer to purchase an integrated solution for simplicity's sake."

But Vast's focus on Supermicro also focuses it on service providers, including those that provide AI functionality in the cloud, McDowell said.

"[Expanding hardware partnerships] is Vast stepping away from the past legacy of hardware appliances," McDowell said.

Adam Armstrong is a TechTarget Editorial news writer covering file and block storage hardware, and private clouds. He previously worked at StorageReview.com.

Dig Deeper on Storage system and application software

Disaster Recovery
Data Backup
Data Center
Sustainability
and ESG
Close