News for hackers. Stuff that matters.
Re:Beating Floating Point at Its Own Game: Posit Arithmetic ... by yorwba on Saturday July 01, 2017 @01:58AM (#14675628)
The LINPACK benchmark feels a bit like bragging. Of course you can demand an exact dot-product operation for any implementation of your standard, and then use it to compute an exact vector-matrix product, but floats could do the same if it were part of the standard. The fact that it requires a 1024-bit accumulator makes me doubt that it would be used for a massively parallel implementation e.g. in GPUs.

The overall idea of choosing a number distribution that loses less precision for common operations and orders of magnitude is pretty compelling, though.

Re:Beating Floating Point at Its Own Game: Posit Arithmetic ... by dnautics on Saturday July 01, 2017 @10:59AM (#14677385)
The accumulator is pipelineable. For most graphics and ml applications you probably don't need the exact dot product. You probably do want it for scientific applications... There is a trade-off in performance, as always, an appeal to calculation speed though falls to the retort of, sure, you can calculate wrong answers with higher throughput if you wished.