This is interesting:
http://research.nvidia.com/sites/de...U:-Multi-Chip-Module-GPUs//p320-Arunkumar.pdf
NVIDIA is assuming a 128 SMs monolithic die is possible on 7 nm, and up to 256 SMs with their MCM solution. In comparison GV100 packs 84 SMs @ 815mm² with TSMC '12 nm FFN'.
...e. For example, ≈ 800mm2 is expected to be the maximum possible die size that can be manufactured [18, 48]. For the purpose of this paper we assume that GPUs with greater than 128 SMs are not manufacturable on a monolithic die.
...in this paper we evaluate building a 256 SM GPU out of four GPMs of 64 SMs each
We show that with these optimizations, a 256 SMs MCM-GPU achieves 45.5% speedup over the largest possible monolithic GPU with 128 SMs. Furthermore, it performs 26.8% better than an equally equipped discrete multi-GPU, and its performance is within 10% of that of a hypothetical monolithic GPU that cannot be built based on today’s technology roadmap.
http://research.nvidia.com/sites/de...U:-Multi-Chip-Module-GPUs//p320-Arunkumar.pdf
NVIDIA is assuming a 128 SMs monolithic die is possible on 7 nm, and up to 256 SMs with their MCM solution. In comparison GV100 packs 84 SMs @ 815mm² with TSMC '12 nm FFN'.