NVIDIA’s first go at making ARM-based processors didn’t go as well as the company had planned. Leaving the failure years behind, the company is now taking a second stab at making ARM-based processors, but instead of going against established names like Intel and AMD in the consumer electronics segment, the company is taking a safer route by targeting large-scale neural network systems for AI data centers – a domain where its GPUs already do the heavy lifting. Called NVIDIA Grace, the fruit of NVIDIA-ARM’s latest joint effort is expected to become commercially available in 2023.
To recall, NVIDIA purchased ARM from SoftBank in a deal valued at $40 billion in September last year. Coming back to NVIDIA Grace, the company says it will offer up to 30X aggregate bandwidth boost compared to today’s fastest servers and ‘up to 10x the performance of today’s fastest servers on the most complex AI and high-performance computing workloads.’ NVIDIA hasn’t namedropped Intel or any of its x86-based Xeon processors here – or AMD’s EPYC range for that matter- but the targets are evident.
Meet NVIDIA Grace – the basic building block of the modern data center. This powerful @Arm-based #CPU allows us to accelerate the largest HPC and #AI workloads. Watch the introduction: https://t.co/z51GV8rkhg #GTC21 pic.twitter.com/EySjYKPsDj
— NVIDIA (@nvidia) April 12, 2021
NVIDIA has gone much into detail about the underlying architecture and core specifics, except for outlining the key areas where it envisions the Grace processor to be used – natural language processing, recommender systems, and AI supercomputing. For now, the company only notes that its latest offering will combine energy-efficient cores with a low-power subsystem. Specifically, NVIDIA Grace will is touted to be the first server CPU to employ the LPDDR5x memory, which the company claims offers 2x bandwidth and 10x higher energy efficiency compared to the DDR4 standard.
The latest NVIDIA offering employs the next-generation Arm Neoverse cores that are said to boost performance as well as energy efficiency. And when paired with NVIDIA’s own GPUs, Grace is said to employ the fourth-gen NVIDIA NVLink interface to offer bidirectional bandwidth of 900 GB/s. Coming to commercial deployment, CSCS and Los Alamos National Laboratory have confirmed plans of making supercomputers based on NVIDIA Grace supercomputers that are expected to go online in 2023.