The NVIDIA GM200 GPU lacks FP64 performance
The NVIDIA Maxwell generation has been with us for quite some time now but we still haven’t seen the greatest chip of this GPU generation – the GM200. The chip will be huge and will offer unmatched performance when it debuts in the upcoming GeForce GTX Titan-X graphics card. Unfortunately no matter how good looking this chip is, it comes with a little dirty secret that we will expose now.
It appears that the GM200 GPU will come with cut-down double-precision (FP64) performance. The reasons for this phenomenon are unclear but the GM200 chip will lack FP64-specific hardware that is mandatory for having acceptable double-precision performance. This is the reason why GM200 will be slow when performing FP64 calculations and as a result the chip may not be featured in professional-class NVIDIA-based graphics cards.
The good news is that gamers will not suffer as the chip is optimized for single-precision (FP32) calculations, which are heavily used in games. Furthermore GM200 appears to be FP32-optimized just like the GM204. And the fact that some hardware is missing means that GM200 will need less electricity and hence it will output less heat when operating.
This information has not been confirmed by NVIDIA yet so there’s still some hope it might be wrong. We will know all about it when the GM200 appears around the March 17-20 time period.