AMD’s MI300 Targets HPC and AI

AMD’s MI300 accelerator will compete with Nvidia's Grace Hopper in the HPC and AI markets. It uses AMD's third-generation CDNA3 architecture and features x86, GPU, and memory die in a single package for improved performance.

Anand Joshi

After challenging Intel in the processor market, AMD’s latest MI300 accelerator is competing with Nvidia’s Grace Hopper in the high-performance computing and AI markets. The MI300 is based on AMD’s third-generation CDNA3 architecture and combines x86, GPU, and a memory die on a single package.

The MI300 is an upgrade to AMD’s current-generation MI250X with dual GPU die. The CDNA3 architecture unifies CPU and GPU memory, sharing the physical memory for improved performance. The MI300 aims to improve training performance by 8× over MI250X and will hit production in the second half of 2023.

AMD targets the HPC market with its MI series GPU products. The MI250X is used in the world’s fastest supercomputer, Frontier, because of its superior vector throughput. The MI300 will be used in the El Capitan supercomputer, slated to replace Frontier when deployed.

In the AI market, AMD has been trying to position its CDNA products as a rival to Nvidia but has had little success because of lackluster AI performance and a weak software stack. AMD expects the MI300 to rival Nvidia in performance, but AMD must still deliver a feature-rich and stable AI software development stack to compete effectively.