Ok, I thought dual GPU graphics cards like the GTX 690 showed up in the Windows device manager as 1 display adapter, but it seems not. I suppose this means it basically just SLI'd the two GPUs together and had similar poor support in games / programs
So, as you say - Hopper would need a complete redesign for MCM (and might not be ready by 2021). Why not simply have 2 powerful GPUs with an 'infinity fabric' type link between them?
Just saw a Nvidia research paper about MCM from 2017 thats kind of interesting. In the conclusion, the authors say "The greatest challenge towards building more powerful GPUs comes from reaching the end of transistor density scaling, combined with the inability to further grow the area of a single monolithic GPU die"
Also, the performance of an MCM GPU could be "within 10% of that of a hypothetical monolithic GPU" (that is too large to build in practice). So, you would probably still lose a bit of performance...
The paper suggests a possible configuration of 4 GPU modules + an interlink (SYS + I/0).
For a single GPU, they estimate that "~800mm2 is expected to be the maximum possible die size that can be manufactured". Interestingly, the A100 7nm Ampere GPU exceeded this, with a die size of 826 mm². I wonder if they already reached the limit for 7nm?
Link: https://research.nvidia.com/sites/default/files/publications/ISCA_2017_MCMGPU.pdf
So, as you say - Hopper would need a complete redesign for MCM (and might not be ready by 2021). Why not simply have 2 powerful GPUs with an 'infinity fabric' type link between them?
Just saw a Nvidia research paper about MCM from 2017 thats kind of interesting. In the conclusion, the authors say "The greatest challenge towards building more powerful GPUs comes from reaching the end of transistor density scaling, combined with the inability to further grow the area of a single monolithic GPU die"
Also, the performance of an MCM GPU could be "within 10% of that of a hypothetical monolithic GPU" (that is too large to build in practice). So, you would probably still lose a bit of performance...
The paper suggests a possible configuration of 4 GPU modules + an interlink (SYS + I/0).
For a single GPU, they estimate that "~800mm2 is expected to be the maximum possible die size that can be manufactured". Interestingly, the A100 7nm Ampere GPU exceeded this, with a die size of 826 mm². I wonder if they already reached the limit for 7nm?
Link: https://research.nvidia.com/sites/default/files/publications/ISCA_2017_MCMGPU.pdf
Last edited: