AMD is expected this year at the introduction of Instinct MI200, accelerator for artificial intelligence calculations intended for supercomputers. The main feature of MI200 is that it will be the first MCM (Multi-Chip Module) project of the US company, that is, we will see two compute dies (GPGPU) based on CDNA 2 architecture on a single package.
The design of microprocessors and GPGPUs based on multiple units connected together through very high bandwidth interconnections represent the near future of the hardware world and we have always seen the first examples thanks to AMD with EPYC server processors and the last two generations of Ryzen desktop (3000/5000).
The MCM project that AMD will adopt on accelerators will also characterize the future generation of Radeon solutions based on RDNA 3 architecture, or at least some high-end models, then going to become a standard that will then evolve again thanks to other design techniques such as 3D stacking. .
A confirmation of this comes from the rumors of the past few hours regarding Instinct MI300, an accelerator expected in 2022 and which should be based on architecture CDNA 3. According to a leaker, the MI300 project could contemplate twice as many die as MI200, so four graphics chiplets called GCD (Graphics Complex Die) on a single package.
At the moment there is no news on the number of total Compute Units aboard MI200 and even less of MI300, even if rumors speak of 128 CU of which only 110 are active, a number which however is not clear if linked to a single GCD (2 x 110 = 220) or both (2 x 55 = 110). All that remains is to wait for new rumors or some official information: the announcement of Instinct MI200 should not be late.