Facebook apparently achieved a performance gain with their designs https://code.fb.com/ai-research/floating-point-math/

For a posit based float with 127 unique unisgned values. If you can tolerate that little precision, you can implement arithmetic as a smallish lookup table.

Tensor cores seem like a fit...

I have not read the paper yet but the abstract says that they compared two hardware adders synthesized from C++ code. It seems like a fair comparison but I wonder whether it might be possible to hand design a much more efficient posit adder that is more competitive with hand designed floating point adders. I can see how the number of lookup tables used in synthesis would be proportional to the number of gates used on a dedicated chip though so maybe there would be no difference in comparing hand optimized adders on ICs.

Bummer. The posit number system (http://www.johngustafson.net/pdfs/BeatingFloatingPoint.pdf) is super-cool, and its inventors often speak of its potential performance advantages. On the other hand, their claims did seems a little "too good to be true" - variable-width fields require mux gates, and somebody needs to pay for those.

Am I interpreting the delay results correctly? That's the time needed for the operation add/multiply to complete?

Because there isn't much difference in that regard... values deviate no more than 25%, which is not insignificant, but not overwhelming either.

What I find even more suprising: Is there almost no difference in latencies for FP-operations between software-implemented and hardware-supported FPs?