Subnormal floating-point numbers are expensive… on Intel processors

I’m curious how this translates to modern GPUs. There was a paper at Oakland a few years ago on using a timing side channel in canvas to leak the contents of things the browser would let you composite but not read back by ensuring that the calculation used subnormals. The paper showed the relative performance of subnormals was an order of magnitude worse relative to normal floating point on GPUs than CPUs.

I’m also curious why the AMD and Graviton FPUs are not slower. Either they decided that subnormals were worth optimising (I wonder what workload that was based on)k or they’re using a different implementation technique that makes subnormals fast.