AMD combines CDNA3 and Zen4 architectures in Instinct MI300 chips for artificial intelligence


AMD Instinct MI300 AI and HPC chips combine the CDNA 3 and Zen4 architectures through advanced packaging; watch

A OMG is investing heavily to develop its solutions Artificial Intelligence and HPC with Instinct MI300 chip. While the MI300X was developed exclusively using the CDNA 3 (Compute DNA for Data Centers) architecture, the MI300A server APU uses advanced packaging to bring Zen4 and CDNA3 cores onto the same die.



You first products with Instinct MI300 chip is expected to be presented at the “Advancing AI” conference on December 6. However, the company has already released some technical details on what to expect.

 

Keeping an eye on the competition

The MI300X AI accelerators will most likely be the focus of AMD’s event. With 8 Graphics Complex Dies (GCDs) interconnected via AMD Infinity Fabric technology, each with eight HBM3 memory modules, the AI ​​GPUs seek to enter the same segment as the Nvidia Hopper H100.

To date, the Nvidia competes with itself on this front, while Intel’s Gaudi 2 chips appear to be trying to meet a most accessible market segment, with slightly lower performance. AMD, for its part, seems to be aiming to bring extremely competitive solutions, to gain space before the arrival of the next generation of Nvidia GPUs for AI.




This is because the MI300X already has configurations above Nvidia’s update for the Hopper generation. AMD GPUs feature 192GB of HBM3 memory, 5.2TB/s of memory bandwidth, compared HBM3e 141GB of 4.8 TB/s from the recently announced Hopper 200.

Additionally, Infinity Fabric brings 896 GB/s of dedicated bandwidth to the interconnectors alone, enabling extremely fast integration between dies.



For entry-level servers, AMD will offer Instinct MI300A chips, first APU designed specifically for AI and HPC. The MI300A’s advanced packaging retains much of the structure of the MI300X, with die integrated via Infinity Fabric interconnectors.

However, two CDNA3 GCDs have been replaced by three Zen4 CCDs (Compute Complex Dis) 8 cores and 16 threads per die. The strategy is similar to that used by Nvidia in the Grace-Hopper superchip, but on a smaller scale, as well as preserving the processor’s x86 architecture.

🛒Buy Nvidia GeForce RTX graphics cards at the best price!

🛒Buy AMD Radeon RX graphics cards at the best price!

Source: Wccftech

Trends on Canaltech:

Source: Terra