For a few years AMD has decided to design two different architectures in the field of GPUs, the well-known RDNA (Radeon DNA) for the consumer graphics sector and CDNA (Compute DNA) for artificial intelligence accelerators to be placed in servers and supercomputers.
After seeing with CDNA 2 and MI200 accelerators the transition to an MCM project for the development of GPUs thanks to the union of two chips on a single package, AMD has even more interesting plans for the future, with MI300 and CDNA 3 architecture. The first thing you can notice when looking at the roadmap is that while consumer roadmaps cover two years, up to 2024, in this case AMD has stopped at 2023.
CDNA 3 will mark AMD’s move in this sector as well 5 nanometers of TSMC. The other piece of information the company shared is that the CDNA 3 architecture will be the cornerstone of one APU per datacenter. Intel will do something similar with Falcon Shores in 2024, so at least a year after AMD, while NVIDIA will join an ARM CPU and a Hopper GPU on the same board with Grace Hopper Superchip from next year.
On a single package AMD will then put one CPU with Zen 4 core, one CDNA 3 GPU, HBM memory and everything you need to create a product capable of offering over 8 times the performance in AI training compared to the current Instinct MI250X accelerator. AMD also indicates a performance improvement per watt in the field of AI more than 5 times by CDNA 3 compared to CDNA 2.
Architecture and production processes aside, from the slides you can see how everything is based on the innovations on the packaging front, with the possibility of installing 3D chiplets capable of accessing a unified memory (cache and HBM) e fully consistent through the fourth generation of what is called Infinity Architecture. Finally, with Instinct MI300 AMD will introduce support for new mathematical formats for an increasingly better and efficient management of loads.