Nvidia considering multichip module design for GPUs

Nvidia researchers have published a paper describing a GPU built as a multichip module. The GPU therefore consists of several parts, which are connected to each other with an interconnect.

Dividing the GPU into multiple modules should make it possible to produce larger GPUs that have more computing power, according to the research. Historically, GPUs have become increasingly powerful as the number of transistors in a chip of the same size continues to increase thanks to smaller manufacturing processes, but in recent years, Moore’s Law has become increasingly difficult to follow.

Moreover, making a single large chip is very expensive. By dividing these into different pieces and making a multi-chip module of, for example, four dies, the costs are reduced. Also in this way a GPU could be made with a size that is impossible as a single monolithic chip.

The researchers describe a GPU that consists of, for example, four GPU modules. In their research, they propose three different architecture optimizations that should ensure a minimum loss in communication between the different modules. According to the researchers, their design approaches the performance of a theoretical monolithic chip of the comparable size. The mcm design would be only 10 percent slower.

It is already possible to combine several GPUs with each other, for example by using two or more video cards in a SLI setup. However, according to the researchers, the mcm-gpu design performs 26.8 percent better than a comparable multi-gpu solution.

Whether and when Nvidia will actually make GPUs based on an MCM design is not yet known. The proposals in the study are all based on theoretical calculations. AMD is already using a multichip module design in its Epyc server processors and in the upcoming Threadripper CPUs for high-end desktops. AMD connects the dies via its Infinity Fabric. AMD has previously indicated that this interconnect can also be used for GPUs.