Did you know in current AI chips,
The energy cost of that "transport" is orders of magnitude higher than the calculation itself. For example, performing a 64-bit floating-point addition might cost roughly 0.1 picojoules (pJ)
But moving that same data from DRAM to the processor can cost 1,000 to 10,000 pJ. We are essentially spending 99% of our energy budget on the "commute."
Something strange appears when you analyze how modern AI systems actually run.
Most of the energy isn’t spent doing math.
It’s spent moving data between memory and compute.
That single observation changes how AI hardware should be built.