haxxiy said:
The improvements in power consumption from smaller nodes are, IPC improvements and different clock-voltage curves aside, a direct function of the physical size of their transistors. The scaling between TSMC's 16FF and N5 is 88 x 70 nm / 51 x 28 nm ) = 4.36 times. For actual GPUs in real-world scenarios, you can see the improvement was more like ~ 3 times between a GeForce 10 and a GeForce 40: So yeah, the overall improvement should be around that ballpark. It has to. Mind that Ampere GPUs can also dual-issue FP32 instructions if there's no simultaneous INT32 going on. So it could claim to be a ~ 4TF console but perform closer to a PS4 than a Series S (without DLSS). |
16nm 🤔