AI Accelerator Chips
New breakthroughs in AI computing were made this year, thanks to new chips from Nvidia, Intel and a handful of startups. In May, Nvidia revealed its new A100 data center GPU, based on its next-generation Ampere architecture, which can accelerate training and inference workloads while also offering the ability to partition itself into as many as seven distinct GPU instances. The company said five DGX A100 systems, each of which include eight A100s, can perform the same amount of training work as 50 previous-generation DGX-1 systems and 600 CPU systems at a tenth of the cost and a twentieth of the power. Near the end of the year, Amazon Web Services launched a new EC2 instance for training deep learning models using Intel’s new Habana Gaudi accelerator, which AWS said could offer up to 40 percent better price-performance than similar cloud instances running Nvidia GPUs. Among the AI chip startups, Graphcore recently said its IPU-M2000 chips can outperform Nvidia’s A100, with some caveats.