It's not too often we get the feeling that some of the technology that we regularly use is reaching its upper-limit, but there comes a time when new ideas need to be pushed forth in order to continue advancements. Take hard drives, for example. It almost seemed impossible to create multi-terabyte drives until PMR (perpendicular magnetic recording) came into play. Today, SMR and helium are helping to take us even further.
On the microprocessor front, the challenges can be just as intense. We wouldn't have 14nm and smaller processors if not for the advancement of transistors, but soon, those die shrinks might not even be enough. NVIDIA clues us into this reality with a just-published white paper, detailing "Multi-Chip-Module" GPUs, and that's just as it sounds.
Remember when we began doubling up on CPU cores in order to provide more overall performance from a single solution? It looks like we're bound to see the same kind of thing for GPUs, except it's not just the cores that are multiplied (there are thousands on the top-end GPUs already), but the heart of the entire GPU, split into separate modules.
NVIDIA calls this design "GPM", standing for GPU Modules, which would integrate into a high-bandwidth bus and integrate power-efficient signaling. As the image above shows, one GPU could be nestled inside of memory chips and system I/O. If a GeForce GTX 1080, for example, were to take on this kind of design today, power and heat would be an inevitable issue, so how that will be tackled should prove interesting. Not that we can assume the GPU Modules laid out here would match current designs, of course; they could wind up being whittled down chips that won't necessarily blow current top-end offerings out of the water. At least not at first.
What this MPM-GPU design will help improve is the efficiency accomplished by having multiple GPUs work together. Today's SLI solution isn't as hot as it used to be, but with this move, it could very well be that it won't be as much of a concern in the future.
If Moore's Law didn't rear its ugly head, it seems likely that NVIDIA could build the best possible GPUs when sticking to current designs, but MCM-GPU is a good alternative. When we'll ever see such a design is up-in-the-air, though. If you're tired of waiting, feel free to hit up the white paper below, as that should keep you busy for quite some time.