I'm aware of fast RNGs, but even compared to HW floating point operations, I believe they're still more expensive than Khan summation. HW circuits maybe could d

I'm aware of fast RNGs, but even compared to HW floating point operations, I believe they're still more expensive than Khan summation. HW circuits maybe could do well. I see that most of the interest is around LLMs and doing quantized sums (gathering from the fact that Intel has shipped this in their accelerator), but this came up in the WiFi positioning code I was working on 10 years ago (we were using "classical" f64).