Semiconductor companies have been exploring ways to step out of the traditional monolithic GPU die design for a while now, searching for something that will enable better performance scaling while keeping production costs at a sane level. Nvidia’s latest approach to driving things forward is the introduction of 3D die stacking using through-silicon via (TSV) technology with an enhanced power delivery approach. This sounds similar to technology we’ve already heard about from AMD, Intel, and TSMC, but there are some differences.
We already know that Nvidia has been planning to move away from monolithic die design. The company has been actively exploring ways of getting more performance using different packaging techniques, the most recent being the use of multi-chip modules (MCM) to build GPUs with continued performance scalability.
Back in 2017, Nvidia presented its MCM-GPU design at the International Symposium on Computer Architecture (ISCA). Nvidia planned to use multiple dies of logic to interconnect the massive amount of cores and develop new GPUs with continuous performance improvement while managing costs. As GPU dies get bigger, their costs increase exponentially, so making a few interconnected smaller dies is a more cost-effective solution. The MCM-GPU packaging approach solves this as it connects multiple dies that provide a huge performance uplift in return.