NVIDIA is one of the biggest players in the computing hardware game, particularly with its GPUs and DPUs. Yet, for a company that focuses so much on data center computing, NVIDIA has never released its own data center-facing CPU—until this week.
NVIDIA Grace is the company’s first data center CPU. Image used courtesy NVIDIA
With what many are calling a direct shot at Intel, NVIDIA announced “Grace,” its first data center CPU. What do Grace’s technical specs offer the industry?
NVIDIA’s First Data Center CPU
While some of the granular technical details are not yet released, we can analyze what we do know about the CPU.
A highly specialized processor, Grace was designed specifically to bolster AI applications with natural language processing and AI supercomputing being amongst the two most pointed applications. The CPU, which is meant to be utilized alongside NVIDIA’s GPUs, is said to be based on Arm Neoverse cores and an “innovative low-power” memory subsystem.
The subsystem is based on LPDDR5x memory, and NVIDIA claims that it will deliver twice the bandwidth and ten times better energy efficiency compared to DDR4 memory. One highlighted feature of the new memory subsystem is unified cache coherence with a single memory address space, resulting in less data movement energy consumed and overall improvements in power efficiency.
NVIDIA claims that its Grace/GPU systems can increase system memory by 30 times over x86 based solutions. Image used courtesy of NVIDIA
Altogether, NVIDIA claims that systems that couple their GPUs with Grace will see ten times improvement from the NVIDIA DGX-based systems, which run on x86 CPUs.
Will Grace Challenge Intel CPUs?
While the CPU won’t be released until 2023, NVIDIA feels confident about the chip’s impact on AI- and HPC-accelerated computing.
NVIDIA’s trajectory for its key GPU (Ampere), CPU (Grace), and DPU (BlueField) families. Image used courtesy of NVIDIA
For starters, this new chip is signaling a new era for NVIDIA – a company historically known for its GPUs. Previous NVIDIA systems leveraged out-of-house CPUs with x86 cores, making this change noteworthy in that it’s both in-house and Arm-based. Apple made a similar move last year when it too moved to an in-house Arm-based CPU as opposed to previous Intel-based solutions.
This change also increases questions about Intel’s position in the marketplace. With NVIDIA’s pending acquisition of Arm, it’s possible that the company may become a direct competitor to Intel’s CPUs.
Governments Tap Grace for Supercomputing
As of now, NVIDIA has some big plans for Grace in the near term.
The U.S. Department of Energy is planning to build a new supercomputer based on the Grace CPU. Beyond that, NVIDIA intends to create its own AI supercomputer to be located in the Swiss Alps, also based on the Grace system. This supercomputer will be used by the Swiss National Computing Center for applications including climate prediction and scientific simulations.