I think they're just referring to 3D rendering. I'm not sure how (or if) traditional memory buses will be feasible on shrinking geometries.
The way I see it you have the problem of needing a very wide bus running at a very fast speed. As you ramp up the clock, the IO will burn more and more power, probably disproportionately compared to the core. With shrinking geometries, the chips themselves will be pad limited as well. I think power wasted in IO will become a significant contributor to the overall power consumption.
If you look at a 7970 it has a bus width of 384 bits, 12 GDDR5 chips, and BW of 264 GB/sec. Let's assume that a 20nm, 9970 will double the performance of the 7970. I think memory BW will probably have to scale accordingly, but I don't think a 512+ bit bus at 7+ Gbps will really be doable (at least for a reasonably priced product, not some vanity card like a 7990)
Long term, it wouldn't surprise me to see all GPU's go with 2.5D stacking of memory. People say that the next gen consoles won't be bleeding edge, I think they will be, just not in raw performance, but the overall system architecture and the efficiency that it can achieve.
The way I see it you have the problem of needing a very wide bus running at a very fast speed. As you ramp up the clock, the IO will burn more and more power, probably disproportionately compared to the core. With shrinking geometries, the chips themselves will be pad limited as well. I think power wasted in IO will become a significant contributor to the overall power consumption.
If you look at a 7970 it has a bus width of 384 bits, 12 GDDR5 chips, and BW of 264 GB/sec. Let's assume that a 20nm, 9970 will double the performance of the 7970. I think memory BW will probably have to scale accordingly, but I don't think a 512+ bit bus at 7+ Gbps will really be doable (at least for a reasonably priced product, not some vanity card like a 7990)
Long term, it wouldn't surprise me to see all GPU's go with 2.5D stacking of memory. People say that the next gen consoles won't be bleeding edge, I think they will be, just not in raw performance, but the overall system architecture and the efficiency that it can achieve.