posits: new format for FP numbers
 

They are described in the IEEE article at the end but briefly, they seem to be mostly used for Neural Network
computations. However, the accuracy increase in test implementations has been eye popping; about 10^4 !!!
All major hardware vendors are interested which bodes well for an affordable implementation in the not-too-distant
future.
Matrix multiplication accuracy increases are specifically mentioned.
Perhaps error control in matmul operations would be trivial or unnecessary !!


[URL]https://spectrum.ieee.org/floating-point-numbers-posits-processor[/URL]