NVIDIA Investigated Multi-Chip-Module GPUs in Research Paper

[Guest_Jim_*]

Necessity is the mother of invention, and the semiconductor industry is quickly approaching a limit to current technologies. This is because consumers and professionals keep demanding more and more performance but there are physical limitations we are approaching. A solution needs to be found for this demand to be met and NVIDIA, Arizona State University, University of Texas at Austin, and Barcelona Supercomputer Center recently worked together to research one of these approaches: multi-chip-modules.

To continue increasing the amount of computing power, more transistors are needed and traditionally this has been achieved by improving monolithic designs or, in the case of GPUs using a multi-GPU system. The monolithic approach has worked for a long time now, but is approaching the limits of silicon feature size and the aperture size of the lithography systems used to create the processors. A multi-GPU system removes the size concern but presents latency issues because of how long it takes to transmit information between two different processors as well as load-balancing concerns. By going with a multi-chip-module GPU though, which stitches together multiple processors onto a single substrate, the size limitations are worked around with far less latency than a multi-GPU system.

According to the research paper, an MCM-GPU appears to very much be a desirable solution, based on the simulations run as part of the work. With an optimized design to address bandwidth, latency, and load-imbalance concerns between GPU modules, an MCM-GPU with 256 streaming multiprocessors (SMs) was 45.5% faster than the largest possible monolithic GPU with 128 SMs and 26.8% better than a multi-GPU system using two such GPUs. When compared to a hypothetical 256 SM monolithic GPU, the MCM-GPU came within 10% of its performance, so while it is not perfect, it is very close.

AMD is also working on MCM designs, with the recently released EPYC CPUs being MCM processors using Infinity Fabric to connect four Zeppelin dies, and Navi GPU architecture likely to use Infinity Fabric in a similar manner. When NVIDIA might actually create an MCM-GPU is hard to guess, but with both companies at least considering this direction, it will be interesting to see what comes.

Source: NVIDIA

Back to original news post