AMD has unveiled its latest Accelerated Processing Unit (APU), built on the Zen 4 architecture in the form of the MI300A – which is expected to give Nvidia a run for its money as organizations buy up components en masse to power AI workloads.
The unit – the first APU for data center, AI and high-performance computing (HPC) workloads – consists of 24 threaded CPU cores alongside 228 CDNA 3 compute units that form the GPU element.
The APU comes alongside unified 128GB of high-bandwidth memory (HBM) in the form of eight stacks of HBM3, split between both CPU and GPU, unlike the third generation of the APU, which had dedicated memory units for each. The components are also included in an interesting design that combines 13 chiplets in a 3.5D packaging method. It is also the largest chip AMD has ever made, with 153 billion transistors. It also has a central 256MB Infinity Cache that maximizes bandwidth and latency for data flowing through the chip.
Stealing the crown from Nvidia
AMD launched its new chip alongside the MI300X AI accelerator, which features eight stacks of HBM3 memory alongside eight 5nm CDNA 3 GPU chiplets, hoping to match or even overtake one of the best GPUs ever made. This forms the MI300 series.
With the Instinct MI300A, however, AMD is hoping the accelerator kicks Nvidia's sense of in-demand chips off their pedestal, including the H100 GPU and the GH200 chip.
AMD test results show up to twice the theoretical peak HPC performance compared to the H100 SMX – four times the performance versus the H100 on some workloads – twice the peak performance per watt compared to the GH200 – as well as AI performance on par or just shy of what Nvidia's H100 can achieve.
AMD's partners so far include HPE, Eviden Gigabyte and Supermicro, but one of the most exciting proposals for the company's latest HPC chip is its inclusion in the El Capitan supercomputer.
This is expected to be the world's first dual-exaflop supercomputer when it becomes operational next year.