3 ms·
I got excited about 8-bit floating point being standardized, which would be actually very useful in machine learning if processors supported it. I really would
by rspeer 10y ago
I got excited about 8-bit floating point being standardized, which would be actually very useful in machine learning if processors supported it. I really would want to be able to use SIMD operations that affect eight 8-bit floats at the same time (without the silly multiplier by 1.25).
Being able to shove more data through is better than each individual value being more accurate.
The fact that this is a joke is cheapening the real-world motivation for minifloats.
- deepnotderp 10y agoYeah, there's a genuine need for it, it's not just a joke. But the problem is that floating point people think that it's a magic cure to accuracy and precision issues, when it's really not (just look at Kahan summation). I'm with Gustafson on this one.
- jacobolus 10y agoIf you want an 8-bit float, Gustafson’s “sigmoid” number idea starts to have some noticeable advantage. https://groups.google.com/forum/#!topic/unum-computing/5LYpAtyP9LY https://groups.google.com/forum/#!topic/unum-computing/5LYpA...