This is it - you can't just compare FP32 TFLOPS. The newer cards have a ton of extra precisions, custom cores for certain workloads, and on chip memory, all of

This is it - you can't just compare FP32 TFLOPS. The newer cards have a ton of extra precisions, custom cores for certain workloads, and on chip memory, all of which use silicon area and transistors, but none of which boost the FP32 TFLOPS metric. I could design you a chip that is nothing but FP32 multipliers and adders that has, theoretically, a ridiculous TFLOPS per mm^2, but it would be next to useless in any real workload.