NVIDIA, VAST Data partner to create AI superclusters
VAST Data has announced it has partnered with NVIDIA to build the foundation for AI superclusters, a next-gen storage platform called Ceres.
VAST explains that Ceres is enabled by the company’s Universal Storage data platform and built using new hardware technologies, including NVIDIA BlueField data processing units (DPUs) and ruler-based hyperscale flash drives for better performance and to simplify serviceability and reduce data center costs.
Additionally, VAST designed Ceres in collaboration with its industry partners to modernise storage for the current AI world and the platform brings with it new levels of speed, resilience, modularity, and data center efficiency.
VAST adds that Ceres will allow the company to equip enterprises and service providers with capabilities previously only afforded to larger hyperscale providers.
“A year ago, we shared our vision for hyperscale data infrastructure to the industry, and we’ve been amazed by the collaboration and support for this vision that has come back from industry partners,” VAST Data co-founder and CMO Jeff Denworth says.
“While explosive data growth continues to overwhelm organisations that are increasingly challenged to find value in vast reserves of data, Ceres enables customers to realise a future of at-scale AI and analytics on all of their data as they build to NVIDIA DGX SuperPOD scale and beyond.”
Ceres provides the following benefits:
- Increased Performance, Better Power and Space Efficiencies. Using new NVIDIA BlueField technology, which combines network and NVMe device management services into a low-power Arm-based SoC, DPUs make it possible to build NVMe enclosures without the need for extensive and power-hungry x86 processors.
- Transitioning NVMe-over-Fabric (NVMeoF) services from x86 servers to NVIDIA BlueField DPUs makes it possible to achieve a 1U form factor capable of delivering over 60GB/s of performance per enclosure for data-intensive applications, all while consuming significantly less power. In addition, VAST’s DASE architecture is positioned to support DPU-based systems by decoupling storage processing from the flash layer.
- Ultra-Dense Flash Capacity Configurations. Ceres evolves beyond the classic SSD form-factor and instead features new ruler-based, high-density SSDs. Ruler-based flash drives will, over time, pack more flash capacity compared to traditional NVMe drives because of their much larger surface area and airflow-friendly design.
- At launch, VAST has partnered with Solidigm to certify its 15TB and 30TB E1.L SSDs (long ruler format), packing up to 675TB of raw flash in only 1U of rack space. With VAST Data’s Similarity-based data reduction algorithms, Ceres can manage nearly 2PB of effective capacity per enclosure at an average 3:1 data reduction ratio.
- Additionally, VAST’s innovative write shaping techniques extend quad level cell (QLC) and penta level cell (PLC) flash endurance, while advanced erasure coding also dramatically shortens the time to rebuild ultra-high-capacity storage devices.
- Simplified Serviceability for Disaggregated Storage Clusters.
VAST says Ceres was also engineered to solve several problems that its customers have faced when dealing with high-density storage systems, including:
- Eliminating the need to slide systems in and out of racks, and the need for cable management, by making the system fully front and rear serviceable
- Reducing the upfront hardware costs with a minimum capacity entry point of 338TB while supporting seamless cluster scaling to hundreds of petabytes
- Improving rack-scale resilience with less hardware required for customers that choose to enable full-enclosure failover in Universal Storage clusters
- Allowing customers to mix and match Ceres with previous generations of VAST-supported hardware to enable the infinite cluster lifecycle
Furthermore, VAST is collaborating with NVIDIA on new storage services to enable zero-trust security and offload functionality with client-side DPUs, including those introduced in the recently announced NVIDIA DGX SuperPOD configurations.
This new platform design will initially be manufactured by VAST design partners such as AIC (for commercial applications) and Mercury Systems (for rugged/defence applications). VAST notes that it will serve as the data capacity building blocks of its Universal Storage clusters.
Moreover, VAST says organisations have already chosen Ceres with some of the world’s largest computing environments, and the data platform company has received software orders to support over 170PBs of data capacity to be deployed on Ceres platforms.
VAST has also announced that it is in the process of certifying Ceres for NVIDIA DGX SuperPOD.
SuperPOD is tailor-made for large-scale AI workloads, utilising high-performance storage and networking to provide customers with a turnkey AI data center solution for enterprises.
“Enterprises around the world are using AI to transform their data into insights and services that boost customer satisfaction, increase operational efficiency and bring new products to market,” NVIDIA DGX systems vice president and general manager Charlie Boyle says.
“NVIDIA DGX SuperPOD and NVIDIA BlueField DPUs, paired with the VAST Ceres platform, offer customers the option to integrate the world’s leading AI infrastructure with high-performance NAS storage certified to meet the demands of advanced AI workloads.”
VAST Data Universal Storage certification for NVIDIA DGX SuperPOD is slated for availability in Q3 2022.