Now, here are the numbers for a reference NVIDIA GPU. Pay specific attention to the change in ratio:

• 6,100 Gflops/sec
• 84 Gwords/sec
• Ratio is 72/1

So, just in terms of raw operations per chunk of memory, GPUs have a clear advantage.

A full detour into microprocessor design is of course outside the scope of this book, but it is useful to think about the processor's distribution of memory and compute units. The design philosophy for modern chips can be summed up as cram as many floating-point units onto the chip as possible to achieve the maximum computation relative to the power required/heat generated.

The idea is to keep those ALUs as full as possible, thus minimizing the amount of time they sit idle while memory gets filled.