AMD was late to AI and seemed like it had been caught far behind Intel and Nvidia. Intel had been buying Nervana for its AI chip in 2016 an Nvidia had been recruiting talent at NeurIPS. At the time AMD was focused on getting back into the server market so it had a few other things to do.   Six years later, AMD has successfully reasserted itself in the data center market, and its AI plans are becoming clearer.

Firstly it bought up Xilinx. Xilinx had already built its Vitus software stack for its Versal Adaptive Compute (ACAP) products which incorporated FPGA fabrics, Arm CPUs, and AI accelerators. In the last year, Xilinx’s former CEO Victor Peng, was named AMD President and tasked with developing AMD’s AI strategy. Changes are beginning to show. AMD now has a broad portfolio that could be used for AI including CPUs, GPUs and Adaptive compute elements.

Now it seems that AMD is starting to spit out AI products. The AMD Data Center & AI Technology Premiere Event was the introduction of a new variant of its Instinct MI300 AI supercomputing hybrid processor designed to support generative AI models.

The company had previously announced MI-300 with three CDNA3 GPU chiplets and one Genoa Zen-4 based Core Complex Die (CCD) chiplet. The initial version of the MI300 has 146B transistors and is designed for the coming two-exaflop El Capitan supercomputer. The chiplet design allow AMD to put more transistors in a package than is possible with a monolithic die and to tightly couple CPUs, GPUs, and high-bandwidth memory (HBM).