AMD is making some serious moves in the AI and high performance computing space, with the launch of the MI300-series accelerators that are poised to take on Nvidia. These chips, known as the Instinct MI300A APUs and MI300X GPUs, are designed to close the performance gap with AMD’s rival and are shaping up to be a game-changer.
When it comes to AI, the performance gap between AMD’s MI250X and Nvidia’s H100 was evident. While the MI250X excelled at precise FP64 mathematics, it fell short on lower precision workloads, where Nvidia’s H100 came out on top. But with the launch of the MI300-series, it looks like AMD has finally taken the lead.
The Instinct MI300X is boasting some impressive numbers. It not only matches but exceeds the speed of Nvidia’s H100, delivering 2.6 petaFLOPS of performance compared to Nvidia’s 1.979 teraFLOPS at FP8. AMD is also claiming a 32 percent higher performance in AI workloads compared to the H100. The MI300A is also promising a significant lead in HPC workloads over the H100.
What makes these new chips truly unique is their architecture and packaging. The MI300X features 12 I/O and GPU dies packed together in a new 2.5D and 3.5D packaging technique. This allows for faster and more efficient communication between the components, resulting in enormous bandwidth measured in terabytes per second.
Overall, AMD’s MI300-series accelerators are poised to be a game-changer in the AI and HPC space, and it looks like Nvidia might have some serious competition on its hands.