In the quest for GPU dominance, we'll see both NVIDIA and AMD shifting away from increasingly larger GPU dies, to many smaller GPU dies. NVIDIA has recently talked about building a MCM package (Multi-Chip-Module package) that is a collection of chips (GPU/CPU/memory/controllers) onto the same chip interposer (similar to GPU/HBM on the interposer of Fiji/Fury X and Vega/RX Vega) that are interconnected through ridiculously fast I/O lanes.
This has already been done, but not on the consumer side of NVIDIA's GeForce graphics cards. AMD is ahead of the game here, as they're using their next-gen Infinity Fabric link between EPYC CPU dies, with 8 cores per die that are interconnected through Infinity Fabric. AMD will be using this same method for their next-gen Navi GPU architecture, which should debut sometime later this year or in 2018.
NVIDIA is proposing that their MCM design might be very different to the traditional design of massive GPU and a couple of DRAM dies on the package, changing it over to MCM and using smaller versions of their GPUs - but just many more of them, as well as much more DRAM dies. The GPU and DRAM dies would be connected through I/O and the controller chipset on-die, versus on-chip.
Better yet, NVIDIA simulated the performance of 256 SMs based on MCM GPUs, with 64 SMs per GPU. Right now, NVIDIA's flagship Volta GPU rocks 84 SMs with 5376 cores... but compared to 256 SM MCM package, we're looking at a mind blowing 16,384 cores... a 3x increase. Using their own in-house GPU simulator, NVIDIA noticed that the MCM-GPU was 45.5% faster than the monolithic GPU, and when compared to multiple GPUs on the same board (like the Radeon R9 295X2 for example), the MCM-GPU was still 26.8% faster.
With both NVIDIA and AMD working on MCM-GPU technology, I can't help but get excited. Is this the next step in GPU technology? Yes. Consider it the same as shifting away from ever increasingly faster single-core CPUs back in the day to the first dual-core CPU, and see where we are now. NVIDIA and AMD will shift to many smaller GPUs on the same interposer, with a mix of GDDR6 and HBM2 memory technologies. To say that I'm excited would be a radical understatement.