NVIDIA Announces AI Storage Breakthrough: GPU Direct Access to Storage Devices Achieves 3.21x Performance Gain

According to NVIDIA, the company announced at this week's Future of Memory and Storage conference multiple AI storage technology advances designed to allow GPUs to directly access storage devices, reducing CPU intermediation and bringing storage access speeds closer to memory speeds.

NVIDIA's Vera CPU, integrated into the Vera BlueField-4 STX platform, achieved a 3.21x data throughput increase compared to traditional x86 CPUs when performing simultaneous data compression and encryption. The company is open-sourcing cuFile API and its underlying storage software architecture, enabling GPUs to directly read from and write to storage devices. NVIDIA also introduced the SCADA (scaled, accelerated data access) framework to optimize data retrieval from large datasets. More than 40 storage and flash memory vendors, including DDN, KIOXIA, and Micron, have joined NVIDIA's Storage-Next initiative, with first maintainers including Google, Intel, NVIDIA, and Meta, to collectively develop next-generation AI storage standards.