I am thinking that multi-chip module (What what we see on Ryzen CPU) is likely the future for GPU. What do you think?
Till now, we have only seen single die GPUs. The only way to have higher performance is to add another graphics card. However, this SLI/CF has other challenges and has received very little support these days.
GPU silicon is getting so big that it easily reaches the limit for a single die.I am thinking maybe future CPU could be multi-chip so we can get around this issue.
https://www.tomshardware.com/news/intel-xe-graphics-all-we-know
Eg, for Intel Graphics. We have slice common and sub slice forming up a slice. So, maybe next time, the slice common is separate (like Ryzen I/O die), then the sub slice are like the chiplets. Eg, a GPU equivalent to RTX2080 may have just 1 slice common and 4 chiplets. 2080ti may have 6 chiplets. It can also scale to 8 chiplets or more.
The main benefit is that we can go beyond the limit of a single die. Also since its a single GPU, we won't have the issues faced by SLI/CF. I also feel cost may be a benefit. Multiple small dies are likely cheaper than a single large one. Yield will also increase.
Main downside will be a slightly drop in scaling performance due to higher latency (same as current Ryzen). Of course, power consumption will go higher as we add more chiplets So, we will probably need liquid cooling (FuryX has shown liquid cooling is possible).
Till now, we have only seen single die GPUs. The only way to have higher performance is to add another graphics card. However, this SLI/CF has other challenges and has received very little support these days.
GPU silicon is getting so big that it easily reaches the limit for a single die.I am thinking maybe future CPU could be multi-chip so we can get around this issue.
https://www.tomshardware.com/news/intel-xe-graphics-all-we-know
Eg, for Intel Graphics. We have slice common and sub slice forming up a slice. So, maybe next time, the slice common is separate (like Ryzen I/O die), then the sub slice are like the chiplets. Eg, a GPU equivalent to RTX2080 may have just 1 slice common and 4 chiplets. 2080ti may have 6 chiplets. It can also scale to 8 chiplets or more.
The main benefit is that we can go beyond the limit of a single die. Also since its a single GPU, we won't have the issues faced by SLI/CF. I also feel cost may be a benefit. Multiple small dies are likely cheaper than a single large one. Yield will also increase.
Main downside will be a slightly drop in scaling performance due to higher latency (same as current Ryzen). Of course, power consumption will go higher as we add more chiplets So, we will probably need liquid cooling (FuryX has shown liquid cooling is possible).
