4 ms·
Indeed, which comes with its own advantages and disadvantages. As someone that has to always care about performance, Machine Learning, I am very happy not to ha
by ninjin 8y ago
Indeed, which comes with its own advantages and disadvantages. As someone that has to always care about performance, Machine Learning, I am very happy not to have to worry about dependencies not being compiled to take advantage of the latest and greatest CPU instructions on a specific box in a very heterogeneous cluster. Now, you could just compile it all on each deployment, but something as large as say TensorFlow takes ages and not everyone gets to enjoy having great devops at your beck and call.
- oblio 8y agoI'm not sure I understand what you mean: you prefer compilation on the target machines or you prefer a static binary?
- ninjin 8y agoCompilation on the target machine, in my case. Performance trumps portability for me, but I can totally see why this is not the case for everyone.
- oblio 8y agoIn theory that can be solved with the best of both worlds: have the runtime and JIT in the final binary. That way you still get easy deployment with 1 big file and high performance after the JIT warms up, which should be fast. I think it's just a problem of manpower on the Julia developer side.
- ninjin 8y agoI wholeheartedly agree, one could also note how this is Java-esque. My hope is that someone far more well-versed than me in this area will find the time and push for it. On that note, one wonderful thing about Julia is how easy it is to contribute to the core language since it is written in Julia. It really is amazing to see people from so many disciplines come together to push a “language for science as a whole”.