3 ms·
I'm torn between two reactions: - 7.5% sounds like a small improvement for a structural change introducing ML profile prediction - a ~250KB model making impro
by abeppu 2y ago
I'm torn between two reactions:
- 7.5% sounds like a small improvement for a structural change introducing ML profile prediction
- a ~250KB model making improvements in an area where humans have been working for a long time seems impressive
- sitkack 2y ago7.5% is huge when most optimizations are notable if they can get above 0.5%, there is NO statistical difference between -O2 and -O3 with Clang. This is a huge result, of which many will follow.
- almostgotcaught 2y ago> 7.5% sounds like a small improvement for a structural change introducing ML profile prediction What world do you live in? There's a cppcon video from 2013 where andrei alexandrescu talks about optimizing the hhvm (hip-hop php compiler) at FB. He mentions that 1% improvement pays for his best engineer for 10 years. I don't know how widely graalvm is deployed but I wouldn't be surprised if it's at similar scale and thus that unimpressive 7.5% translates into ~75MM in savings.