In tests of real workloads, the Nvidia Grace CPU Superchip scored 2x performance gains over x86 processors at the same power envelope across major data centre CPU applications.
This means data centers can handle twice as much peak traffic; they can cut their power bills by as much as half; and they can pack more punch into the confined spaces at the edge of their networks.
Data centre managers need these options to thrive in today’s energy-efficient era. Physics no longer lets engineers pack more transistors in the same space at the same power, which is why new CPUs offer gains over prior generations of less than 30%, and why a growing number of data centres are power capped.
Compute demand is growing 10% a year in the US, and will double in the eight years from 2022-2030, according to a McKinsey study.
“Pressure to make data centres sustainable is therefore high, and some regulators and governments are imposing sustainability standards on newly built data centres,” it says.
The Grace CPU delivers efficient performance with its ultra-fast fabric that connects 72 Arm Neoverse V2 cores in a single die that sports 3,2Tb per second in fabric bisection bandwidth, a standard measure of throughput. It connects two of those dies in a superchip package with the Nvidia NVLink-C2C interconnect, delivering 900 GB/s of bandwidth. Finally, it’s the first data centre CPU to use server-class LPDDR5X memory, that provides up to 50% more memory bandwidth at similar cost but one-eighth the power of typical server memory. And its compact size enables 2x the density of typical card-based memory designs.
Nvidia engineers running real data centre workloads on Grace found that compared to other CPUs in data centers using the same power footprint, Grace is 2,3x faster for microservices; 2x faster in memory-intensive data processing; and 1,9 x faster in computational fluid dynamics, used in many technical computing apps.