Informal testing on one machine shows 2.5x viz BigFloat at worst over all sizes.
I am not changing the names (these are not submissions to the package ecosystem -- they are available), and I recognize that these are not IEEE754-2008 Binary128 .. format. I have no idea about how to compare Arb and Unum. On Saturday, January 2, 2016 at 11:50:58 AM UTC-5, Scott Jones wrote: > > This is very interesting! I'm curious as to how it will compare to Unums, > as it seems both Fredrik Johannson's Arb and Unums are trying to fix a lot > of the same problems with current floating point. > > I'm curious about the Float128 type - this is based on Arb, and doesn't > seem to match the IEEE 754 binary128 floating point standard (which has 112 > bits of fraction (plus "hidden" 1-bit), 15 bits of exponent + 1 bit sign). > Should these maybe be called something else, so as not to cause confusion > with the IEEE standard binary floating point types? > > Do you have any benchmarks comparing these to BigFloats with precision set > to 128, 256, 512, 1024 bits? > > Scott > > On Thursday, December 31, 2015 at 2:50:55 AM UTC-5, Jeffrey Sarnoff wrote: >> >> Responding to a suggestion, these types also are available in separate >> modules: Floats128.jl <https://github.com/J-Sarnoff/Floats128.jl>, >> Floats256.jl <https://github.com/J-Sarnoff/Floats256.jl>, Floats512.jl >> <https://github.com/J-Sarnoff/Floats512.jl>, Floats1024.jl >> <https://github.com/J-Sarnoff/Floats1024.jl>. >> >
