Let's talk about CUDA cores. So it looks like that seemingly massively increased number of shaders isn't the true story and neither are the TFLOPS. It has been noticed that performance of the new cards does not scale linearly with the core count as it usually does.
So the deal is that Nvidia basically invented hyperthreading for shaders and is selling it as double the shader count, which I find incredibly misleading. Two calculations per clock in the same shader just doesn't scale as well as 2 separate shaders. Yet they also use that "doubled" shader count for the calculation of TFLOPS. That means in real world performance Nvidia's shader count and TFLOPS are now worth less than they were with Turing and probably even below AMD.
But there is another theory I have I'd like some input on.
I believe that possibly applications are not yet able to fully utilize the massively increased logical shader count as you can parallelize only so much. Which is why I believe that performance on Ampere and any card that uses the new shaders will slowly increase to close the efficiency gap over the next 5-10 years.
If you demand respect or gratitude for your volunteer work, you're doing volunteering wrong.
