5 ms·

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 a

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?..

ipsum2 6y ago
ARM Neon instructions are pretty fast for math, but I haven't looked for benchmarks: http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batched-BLAS-2017/talk18-goodyer.pdf http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batc...
- eslaught 6y ago
  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.
lasagnaphil 6y ago
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.
mantap 6y ago
PCs are for gaming. PCs are for work. Macs are for... students and middle managers apparently.
- MathCodeLove 6y ago
  And iOS development
- ashtonkem 6y ago
  And about 30% of software devs, according to Stack Overflow.
- chpmrc 6y ago
  Right, apparently only a quarter of all developers uses them (https://insights.stackoverflow.com/survey/2019#technology-_-developers-primary-operating-systems https://insights.stackoverflow.com/survey/2019#technology-_-...).
  - mantap 6y ago
    Including me. Will those same developers buy a machine that is completely locked down? I will not.
    - chpmrc 6y ago
      Sorry, I don't understand what you're referring to. Locked down in what way?
  - fireattack 6y ago
    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).
YetAnotherNick 6y ago
I think for this reason they are not going to discontinue intel based macs as x86 is the default for many high performance software.
- Gaelan 6y ago
  They claimed they would "complete this transition" in 2 years.
  - vosper 6y ago
    They also stated they'll be supporting Intel Macs for a long time, and have new, unreleased, Intel-based machines in the pipeline.
kakali 6y ago
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.
messe 6y ago
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.
NathanielLovin 6y ago
numpy/scipy/etc will have to support Accelerate instead of MKL.
- physicsguy 6y ago
  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.
truth_seeker 6y ago
Rosetta 2 will take care of translation and JITing
fluffything 6y ago
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/ 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/ https://www.top500.org/lists/green500/2020/06/), uses ARM CPUs and has an implementation of these for them.
_venkatasg 6y ago
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.
- aldanor 6y ago
  Interesting - could you share a link to the video?
  - ArgyleSound 6y ago
    https://twitter.com/markvillacampa/status/1275200446764912643?s=21 https://twitter.com/markvillacampa/status/127520044676491264... https://developer.apple.com/wwdc20/102 https://developer.apple.com/wwdc20/102

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?..

ipsum2 6y ago

ARM Neon instructions are pretty fast for math, but I haven't looked for benchmarks: http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batched-BLAS-2017/talk18-goodyer.pdf http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batc...

eslaught 6y ago
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.

lasagnaphil 6y ago

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.

mantap 6y ago

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

MathCodeLove 6y ago
And iOS development
ashtonkem 6y ago
And about 30% of software devs, according to Stack Overflow.
chpmrc 6y ago
Right, apparently only a quarter of all developers uses them (https://insights.stackoverflow.com/survey/2019#technology-_-developers-primary-operating-systems https://insights.stackoverflow.com/survey/2019#technology-_-...).
- mantap 6y ago
  Including me. Will those same developers buy a machine that is completely locked down? I will not.
  - chpmrc 6y ago
    Sorry, I don't understand what you're referring to. Locked down in what way?
- fireattack 6y ago
  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).

YetAnotherNick 6y ago

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

Gaelan 6y ago
They claimed they would "complete this transition" in 2 years.
- vosper 6y ago
  They also stated they'll be supporting Intel Macs for a long time, and have new, unreleased, Intel-based machines in the pipeline.

kakali 6y ago

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.

messe 6y ago

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.

NathanielLovin 6y ago

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

physicsguy 6y ago
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.

truth_seeker 6y ago

Rosetta 2 will take care of translation and JITing

fluffything 6y ago

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/ 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/ https://www.top500.org/lists/green500/2020/06/), uses ARM CPUs and has an implementation of these for them.

_venkatasg 6y ago

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.

aldanor 6y ago
Interesting - could you share a link to the video?
- ArgyleSound 6y ago
  https://twitter.com/markvillacampa/status/1275200446764912643?s=21 https://twitter.com/markvillacampa/status/127520044676491264... https://developer.apple.com/wwdc20/102 https://developer.apple.com/wwdc20/102