The Red team is very serious about implementing Multi-Chip-Module design, which we saw on its mainstream CPUs, to GPUs in the coming days. This will significantly increase performance as we will see more GPU cores thanks to the new approach.
This is the reason why we are hearing rumours that AMD’s next GPU architecture will give tremendous performance gains over RDNA 2, which is already very impressive. Now, interestingly, AMD confirmed in its Q2 earnings presentations that “Initial shipments of next-generation AMD Instinct accelerators featuring 2nd Gen CDNA architecture”.
This confirms that Instinct MI200, which is a successor to the Arcturus-based MI100 accelerator, is already being shipped to customers. Previously, it was told that AMD Instinct MI200 CDNA2 ‘Aldebaran’ MCM HPC GPU Accelerator will launch in late 2021.
This Aldebaran GPU will be a dual-die processor and as its being told that it is getting shipped means AMD did end up becoming become the first ever company to ship its next-generation MCM accelerator to customers. This is even before Intel released its Xe-HP(C) accelerator and before NVIDIA launched its H100 accelerator based on Hopper MCM GPU.
We exactly don’t know the specs of this MI200 accelerator GPU but it was unveiled in the previous Linux patch that the AMD Instinct MI200 ‘Aldebaran’ GPU is definitely going to offer HBM2E memory support. HBM2E offers up to 16 GB memory capacity per stack.
The GPU features a 16 KB of L1 cache per CU which makes up 2 MB of the total L1 cache considering 128 Compute Units will be packed by the GPU. So, if each unit was to retain 128 cores, thenthat would sum up to a total core count of 16384.
While MCM is getting implemented to server oriented setups first but it will be interesting to how MCM design impacts or boost the performance that upcoming RDNA 3 is promising next year.