Ever the showman, Nvidia CEO Jensen Huang unveiled the chipmaker’s new A100 data center GPU running Ampere architecture during a video kitchen keynote from his home in California.
“Ladies and gentlemen, the world’s largest graphics card,” Huang said as he struggled to lift a 50-pound graphics card that runs eight of the new GPUs from a lower oven rack. With 1 million drill holes and 30,000 components, he panted, “It is just a technology marvel.”
Huang, who founded Nvidia in 1993 and has always been its CEO, usually crams his keynotes with tons of insights and new products, but this year’s GTC 2020 keynote was a rare sight. More than 46,000 developers registered to see the keynote online because of COVID-19 and were able to witness Huang discuss robotics and AI, alongside data center innovations that have already attracted Argonne National Lab researchers that are studying the novel coronavirus’ spread and future vaccines.
Huang’s oven spectacle occurs at minute 7 in the sixth segment of seven video segments that launched on YouTube on Thursday.
The sixth segment runs 23:45 minutes, the longest of the seven, and is arguably the most significant announcement because of the power of the A100 GPU. It also includes a video action blow-up of the DGX A100 system, which is being used by Argonne. The video was created with Nvidia’s ray tracing graphics technology.
The entire series of kitchen keynote videos runs for 88:22 minutes with the first 12:55 installment here.
Huang saved the A100 for nearly the last in his keynote with good reason. Rumors of the chip’s Ampere architecture have been floated for months, but central to its capabilities is a process called MIG (multi-instance GPU) that allows a company or research firm to run up to seven instances of AI inference work per chip, or 56 such instances in the graphics card that Huang pulled out of the oven. With different instances, companies can turn up AI work on demand as needed or run all together for powerful data crunching. “MIG is going to have a profound impact on how we architect data centers,” Huang said.
Ian Buck, general manager of acclerated computing at Nvidia, said data center operators will be able to activate one or several instances of the A100 with a simple command line interface.
By the numbers, the A100 has 20 times the performance over its predecessors and will act as a machine learning accelerator for data analytics, training and inference in various data analytics, scientific computing and cloud graphics applications. The A100 has 54 billion transistors, the world’s largest chip built on a 7-nm process. It is fabricated by TSMC on a chip on wafer on substrate process (CoWoS). It delivers 1.5 terabytes per second of bandwidth.
Nvidia said 18 cloud service providers and computing system builders expect to incorporate the A100 GPUs, such as Alibaba Cloud, AWS, Baidu Cloud, Cisco, Dell, Google Cloud, Hewlett Packard Enterprise, Microsoft Azure, Oracle and Tencent Cloud.