IgorsLAB published Does NVIDIA for Ampere need a new power supply solution? The current Turing design as a bottleneck for Tensor and AI performance A quote from the article:

I had already reported a few things about NVIDIA’s new generation of Ampere cards in several articles and read the feedback on them. Interesting there were especially the extrapolations of the number of possible CUDA cores, where all comparisons to Turing regarding performance and efficiency and as a consequence also with the possible power consumption were made. Or better: you believe you can attach it to it. However, the sole scaling of the CUDA cores as the scale of an estimateis a real problem, which I would like to explain.

  Does NVIDIA for Ampere need a new power supply solution? The current Turing design as a bottleneck for Tensor and AI performance