Nvidia just unveiled DGX A100, an AI chip with a 5-petaFLOPS performance, roughly equivalent to a 20x jump over previous generations. According to Nvidia CEO Jensen Huang, which announced the chip from his kitchen in California, Nvidia DGX A100 is “the ultimate instrument for advancing AI” as well as “the world’s most advanced system for all AI workloads“.
DGX A100 is a data center-first computing unit. In other words, it is designed for applications running on entire data centers, not on commercial units. According to Nvidia CEO, in 10 years data centers will replace servers as the fundamental unit of computing. By Huang’s words, a single rack of five DGX A100 systems can replace an entire AI training and inference-focused data center, at 1/10th the cost, while using 1/20th the power and 1/25th the space. Nvidia says that a “typical cloud cluster” featuring DGX-1 nodes combined with 600 separate CPUs for inference training could be replaced by five DGX A100, condensing the hardware down from 25 racks to one, the power budget from 630kW to 28kW, and the cost from $11 million to $1 million.
DGX A100 is built upon the new Nvidia’s Ampere GPU architecture, a massive leap forward from previous AI chips in terms of GPU resources. It features eight A100 GPUs, with 40GiB GPU memory apiece. The chip mounts 54 billion transistors, and with a 5-petaFLOPS performance it represents a 20 times increase over the Volta-powered DGX-2 deep learning system. In addition, multiple smaller workloads can be accelerated by partitioning the DGX A100 into up to 56 instances per system, using the A100 multi-instance GPU feature.
“You get all of the overhead of additional memory, CPUs, and power supplies of 56 servers collapsed into one,” Huang said. “The economic value proposition is really off the charts, and that’s the thing that is really exciting.”
Power is not the only thing that takes you by surprise. Indeed, Nvidia is ditching Intel in favour of AMD. Such a move should not take you unaware, given that AMD’s Epyc was already giving Intel a run for their money. Sure, AMD is a competitor for Nvidia. But as far as it seems, Nvidia is willing to compete with AMD if that undermines Intel’s once impregnable lead on chip manufacturing. Also, DGX A100 caters to data centers, a domain where AMD does not represent a threat to Nvidia. Last but not least, Intel is currently more focused on commercial chips, such as Intel 10th Gen Comet Lake-S.