These are solutions for data centers and AI
AMD introduced the monstrous specialized chips Instinct MI300A and Instinct MI300X back in June, but only now it announced the availability of these products, and at the same time revealed more details.
Recall that the Instinct MI300X is a GPU-based accelerator, and the Instinct MI300A — this is actually a hybrid processor, only with incredible characteristics.
The Instinct MI300X consists of 153 billion transistors, making it the largest and most complex GPU in the world. The graphics core has 304 CU units, that is, 19,546 stream processors, and 1216 Matrix Cores. The accelerator also has a huge amount of very fast memory. This is 192 GB HBM3 with a throughput of 5.3 TB/s. For comparison, the Nvidia H100 has up to 80 GB of memory with a bandwidth of 3.35 TB/s.
Actually, a comparison of MI300X and H100 can be seen on one of AMD’s slides. This comparison is not at all in favor of Nvidia’s solution, and the difference is sometimes more than twofold.
Instinct MI300A — It's an APU, but of monstrous size. It consists of 146 billion transistors and contains 24 Zen 4 cores, a graphics core based on CDNA 3 architecture with 228 CU units or 14,592 stream processors, 128 GB of HBM3 memory with a bandwidth of 5.3 TB/s.
Both new chips are manufactured using 5 and 6 nm process technology.