2070 Super beating 3060 Ti on average

[CaptainHalon]

They're pretty close in PPD, but for some reason the 2070 Super seems to edge a bit more out. Anyone know why when Nvidia said the 3060Ti is actually supposed to outperform a 2080 Super?

[JimboPalmer]

Anyone know why when Nvidia said the 3060Ti is actually supposed to outperform a 2080 Super?

I suspect Nvidia did not test F@H.

I bet the new card is better at raytracing or screen smoothing or some other Gaming feature. (Gaming is their target market)

Frames Per Second is not the same as Points Per Day.

[bruce]

JimboPalmer explained it very well. "Performance" is defined by the sales department, depending on the audience being addressed. FAH_performace depends mostly on pure 32-bit floating point performance and doesn't care at all about tensor cores or rasterization processing.

TU104: The 2070Super theoretical FP32 performance is 9.062 TFLOPS, achieved by 2560 shaders @ 1605 MHz
GA104: The 3060 TI theoretical FP32 performance is 16.20 TFLOPS, achieved by 4864 shaders @ 1410 MHz

FP32 performance is not the ONLY factor for FAH performance, but I consider it the most important single number. It can be altered in two ways, either by adding more shaders or by increasing the clock rate. Small proteins do NOT benefit from having more shaders; large proteins do. Those small proteins benefit more from increased clock rates, so it's the 1605 number that matters for them. Only the very largest proteins would be able to fully utilize the 4864 number.

This is a gross oversimplification, of course, but it does get an important point across.

[CaptainHalon]

Small proteins do NOT benefit from having more shaders; large proteins do. Those small proteins benefit more from increased clock rates, so it's the 1605 number that matters for them. Only the very largest proteins would be able to fully utilize the 4864 number.

This makes sense. Thanks.