11-26-2015, 06:54 AM
See what I mean:
![[Image: NVIDIA-We-are-here.jpg]](http://cdn.wccftech.com/wp-content/uploads/2015/11/NVIDIA-We-are-here.jpg)
Meaning - with Pascal, we're seeing 7-14x the GFLOPs/W over what Nvidia had in 2013 (Kepler, although it was still an extremely dumbed-down architecture, DP-wise). But then fast forward 6 years later to 2023, and we're seeing only maybe 2-3x increase.
![[Image: NVIDIA-Pascal-GPU_Roadmap.jpg]](http://cdn.wccftech.com/wp-content/uploads/2015/09/NVIDIA-Pascal-GPU_Roadmap.jpg)
That is, about 70% increase for Volta (per watt), along with 7 DP TFLOPs compared to Pascal's 4 DP TFLOPs. SGEMM is Single Precision General Matrix Multiplication - so it's more meaningful for games, at least when it comes to power efficiency (without assuming that the increase in DP FLOPs output also means that the chip is getting upwards of 70% increase in SP FLOPs output or even overall efficiency. Still, Volta will sport only 1.2TB/s bandwidth HBM2 compared to Pascal's 1TB/s.
Nonetheless, with the first image showing DP performance per watt, whatever comes out in 2023 would probably be as low as 1.5x as efficient as Volta in 2018, assuming that Volta is receiving a similar gain in DP efficiency as with SP efficiency over Pascal.
This is scary, and I'm sure NV would have to work extra hard at ensuring that there is a new memory architecture to replace the projected 120W consumption needed for HBM2 memory with 2TB/s bandwidth (or 160W 1.5x efficient HBM2 with 4TB/s bandwidth - not expected in time for Volta after Pascal).
Meaning - with Pascal, we're seeing 7-14x the GFLOPs/W over what Nvidia had in 2013 (Kepler, although it was still an extremely dumbed-down architecture, DP-wise). But then fast forward 6 years later to 2023, and we're seeing only maybe 2-3x increase.
That is, about 70% increase for Volta (per watt), along with 7 DP TFLOPs compared to Pascal's 4 DP TFLOPs. SGEMM is Single Precision General Matrix Multiplication - so it's more meaningful for games, at least when it comes to power efficiency (without assuming that the increase in DP FLOPs output also means that the chip is getting upwards of 70% increase in SP FLOPs output or even overall efficiency. Still, Volta will sport only 1.2TB/s bandwidth HBM2 compared to Pascal's 1TB/s.
Nonetheless, with the first image showing DP performance per watt, whatever comes out in 2023 would probably be as low as 1.5x as efficient as Volta in 2018, assuming that Volta is receiving a similar gain in DP efficiency as with SP efficiency over Pascal.
This is scary, and I'm sure NV would have to work extra hard at ensuring that there is a new memory architecture to replace the projected 120W consumption needed for HBM2 memory with 2TB/s bandwidth (or 160W 1.5x efficient HBM2 with 4TB/s bandwidth - not expected in time for Volta after Pascal).