Hacker News Reader

Comment by aldanor | Hacker News Reader

So, if I install numpy via conda on a Mac now, it's backed by Intel MKL and is thus amazingly fast. What will it be replaced with? Has anyone at Apple thought about use cases like this?..

View on HN

I bet that they already have some low-level math library that uses ARM NEON intrinsics; you would definitely need them to port performance-demanding apps like Final Cut Pro / Maya / Photoshop / etc.

There are many implementations of this for ARM.

In fact, the fastest (#1 Top500 as of June 2020 https://www.top500.org/lists/top500/2020/06/) super computer in the world, and one of the most efficient (#9 Green500 as of June 2020 https://www.top500.org/lists/green500/2020/06/), uses ARM CPUs and has an implementation of these for them.

numpy is backed by BLAS and LAPACK. It just happens that on your system on macOS those libraries are provided my MKL. There are other implementations of BLAS and LAPACK out there.

In the Platforms State of the Union, they specifically listed Numpy as one of the open source projects that they built for Apple Silicon (along with Python 3 and others). Go to 20 minutes and 35 seconds on that video.

Interesting - could you share a link to the video?

https://twitter.com/markvillacampa/status/127520044676491264...

https://developer.apple.com/wwdc20/102

ARM Neon instructions are pretty fast for math, but I haven't looked for benchmarks:

http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batc...

This isn't just about hardware. It's about software, and an investment in making libraries that are fully optimized for the hardware. The problem here is the long tail of needs is very, very long.

Isn't that what VecLib is for? Also there are other math libraries to use on ARM. I think we use EigenBLAS in our products.

Rosetta 2 will take care of translation and JITing

I think for this reason they are not going to discontinue intel based macs as x86 is the default for many high performance software.

They claimed they would "complete this transition" in 2 years.

They also stated they'll be supporting Intel Macs for a long time, and have new, unreleased, Intel-based machines in the pipeline.

numpy/scipy/etc will have to support Accelerate instead of MKL.

SciPy did support Accelerate, but they dropped it in 2018 because Apple’s implementation of LAPACK was so out of date. Apple have been crap at updating this stuff which is why people don’t use it.

PCs are for gaming. PCs are for work. Macs are for... students and middle managers apparently.

And iOS development

And about 30% of software devs, according to Stack Overflow.

Right, apparently only a quarter of all developers uses them (https://insights.stackoverflow.com/survey/2019#technology-_-...).

Including me. Will those same developers buy a machine that is completely locked down? I will not.

Sorry, I don't understand what you're referring to. Locked down in what way?

I'm surprised that Windows has ~50%.

I'm not a developer myself, but When I engaging with them on various OSS projects I always feel Windows is being treated as a second class citizen in almost every aspect (smaller projects often has no test, tool chain setup, tutorial, or all of above for Windows, as a starter).

Reply on news.ycombinator.com