Here at its Advancing AI 2024 keynote, AMD officially pulled back the veil on several of its commercial and HPC products. Chief among them are the new fifth generation EPYC processor, formerly known as Turin, and new Instinct AI accelerator.
Starting with its latest EPYC CPU, the new product, basically comes with all the modern trimmings that AMD has. That means you’re looking at an enterprise and AI-based processor built around its Zen5 architecture that is, in turn, built using TSMC’s 3nm and 4nm process nodes. Moving on, the lineup also gets up to 192-cores and up to 384-threads on a single die, with operating speeds of up to 5GHz and full AVX512 support.
Compared to its predecessor, AMD’s 5th generation EPYC processor provides a 17% IPC uplift. Lastly, it runs on the SP5 platform which is also compatible with Genoa.
Moving on, there is the AMD Instinct GPU Accelerators. Now, before we continue, we should make clear that, unlike consumer products, you’re not going to be purchasing these products. Not unless you’re planning on setting up LLM machine learning systems or setting up virtual machines (VM).
Getting back on point, AMD’s latest Instinct AI accelerator, the MI325X, and is the follow-up to the chipmaker’s MI300 Series of GPU accelerators. Specs-wise, it runs on the company’s CDNA3 architecture, 256GB HBM3E memory, houses 158 billion transistors, and has a memory bandwidth of 6TB/s.
In terms of performance, AMD promises up to 2.6PF on FP8 performance, and up to 1.3PF for FP16 performance. And that’s just on a single chip.
In true datacentre and HPC fashion, AMD will be shipping out a platform comprising eight MI325X accelerators. These PCBs come with up to 2TB HBM3E memory, 896GB/s Infinity Fabric Bandwidth, and up to 48TB/s memory bandwidth. Performance-wise, there is 20.8PF on FP8 performance and 10.4PF on FP16 performance.
The AMD Instinct MI325X is expected to start shipping out to enterprise customers by the fourth quarter of this year.
