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

> 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.

7.5% 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.

Anyone with a link to the paper?

https://www.sciencedirect.com/science/article/abs/pii/S01641...

Not spending $55.20 on it.

Summary

https://summarize.beloga.xyz/https%3A%2F%2Fmedium.com%2Fgraa...