Summary
This whitepaper provides a detailed examination of the NVIDIA Grace CPU Superchip, which is designed for accelerated computing in modern data centers. It outlines the key features of the Grace CPU, emphasizing its architecture that integrates performance typically found in dual-socket x86-64 platforms into a single superchip. The document describes the high-performance architecture, including the use of NVIDIA's NVLink Chip-2-Chip interconnect, which facilitates a direct connection between chips to alleviate data movement bottlenecks. The Grace CPU Superchip is built with 144 Arm Neoverse V2 cores and offers up to 1TB/s of memory bandwidth within a 500W power envelope. Additionally, the whitepaper discusses the LPDDR5X memory subsystem, which balances bandwidth, energy efficiency, and capacity for large-scale AI and HPC workloads. It also covers the CPU's support for various I/O connectivity options and the innovative features that enhance performance, making it suitable for demanding cloud workloads and enterprise computing.
