It’s this:
Not sure how you’d achieve what we saw without this.
I’d don’t think that PCI express will be the bottleneck here. Perhaps the compression can be more aggressive since it’s the GPU that does the heavy lifting now. My bet here is that even if you can have a sustained read of 500MB/s per second, it may just not be enough. I really want to believe that the 3GB/s of the current NVMes is enough for the higher-end stuff, though.