Why has AMD chosen an MCM design in its new graphics accelerators?

In the end all the leaks were confirmed. AMD has used an MCM design in its new Instinct MI200 graphics accelerators, based on the CDNA 2 architecture. This is important because it means, in short, the abandonment of classic monolithic core designs, although we must bear in mind that, for the moment, it is most likely that we have a stage ahead in which both designs will coexist for several years.

In the case of the Instinct MI200, AMD has used the same concept that we saw in the processors, Ryzen, combining several chips to create a superchip, that is, a superGPU, which will work as a high-performance graphics accelerator. The top-of-the-range Instinct MI200 adds 14,080 shaders, distributed in two blocks of 7,040 shaders, which means that it uses a total of two chips with 110 computing units each. Those two chips have 64GB of HBM2E memory each, and they are interconnected to work as if they were a single superchip.

We already knew all this, but I wanted to remind you to avoid any doubt, since it is key to understanding the explanation that we are going to give you below of why AMD has used this approach. The main reason is wafer-level efficiency. It is not difficult to understand, think how complicated it would be to move a single chip with 14,080 active shaders to the wafer, the chances that something will go wrong, and that it is not possible to get a good number of working chips, is higher than if we bring a design with 7,040 active shaders to the wafer. The latter will have a higher success rate.

From what we have seen in the previous paragraph, more cost efficient on the wafer to make two smaller, simpler chips with 7,040 shaders and combine them to create a 14,080 shader solution that addresses this GPU under a monolithic core design. On the other hand, we must also bear in mind that, although this is the main reason, the cost ratio and the success rate in the wafer, we must not forget that a design of this type has important advantages in terms of scalability.

We already saw it with the Ryzen processors, AMD only had to join chiplets to get processors between 4 and 64 cores, all under a single architecture, which greatly reduced the costs that, previously, had to be faced when developing different architectures for the sectors consumer, notebook and professional sector, and also those derived from the design of monolithic processors with a high number of cores. Cost, wafer efficiency, and scalability, those are, in short, the three keys why AMD has opted for that MCM design in the Instinct MI200.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *