pythonnumpytensorflowanacondaapple-m1

Why is numpy native on M1 Max greatly slower than on old Intel i5?

I just got my new MacBook Pro with M1 Max chip and am setting up Python. I've tried several combinational settings to test speed - now I'm quite confused. First put my questions here:

Evidence supporting my questions is as follows:

Here are the settings I've tried:

1. Python installed by

2. Numpy installed by

conda install -c apple tensorflow-deps
python -m pip install tensorflow-macos
python -m pip install tensorflow-metal

3. Run from

Here is the test code:

import time
import numpy as np
np.random.seed(42)
a = np.random.uniform(size=(300, 300))
runtimes = 10

timecosts = []
for _ in range(runtimes):
    s_time = time.time()
    for i in range(100):
        a += 1
        np.linalg.svd(a)
    timecosts.append(time.time() - s_time)

print(f'mean of {runtimes} runs: {np.mean(timecosts):.5f}s')

and here are the results:

+-----------------------------------+-----------------------+--------------------+
|   Python installed by (run on)→   | Miniforge (native M1) | Anaconda (Rosseta) |
+----------------------+------------+------------+----------+----------+---------+
| Numpy installed by ↓ | Run from → |  Terminal  |  PyCharm | Terminal | PyCharm |
+----------------------+------------+------------+----------+----------+---------+
|          Apple Tensorflow         |   4.19151  |  4.86248 |     /    |    /    |
+-----------------------------------+------------+----------+----------+---------+
|        conda install numpy        |   4.29386  |  4.98370 |  4.10029 | 4.99271 |
+-----------------------------------+------------+----------+----------+---------+

This is quite slow. For comparison,

Here is the CPU information details:

$ sysctl -a | grep -e brand_string -e cpu.core_count
machdep.cpu.brand_string: Intel(R) Core(TM) i5-6360U CPU @ 2.00GHz
machdep.cpu.core_count: 2

% sysctl -a | grep -e brand_string -e cpu.core_count
machdep.cpu.brand_string: Apple M1 Max
machdep.cpu.core_count: 10

I follow instructions strictly from tutorials - but why would all these happen? Is it because of my installation flaws, or because of M1 Max chip? Since my work relies heavily on local runs, local speed is very important to me. Any suggestions to possible solution, or any data points on your own device would be greatly appreciated :)

Solution

  • Update Mar 28 2022: Please see @AndrejHribernik's comment below.

    How to install numpy on M1 Max, with the most accelerated performance (Apple's vecLib)? Here's the answer as of Dec 6 2021.

    Steps

    I. Install miniforge

    So that your Python is run natively on arm64, not translated via Rosseta.

    1. Download Miniforge3-MacOSX-arm64.sh, then
    2. Run the script, then open another shell
    $ bash Miniforge3-MacOSX-arm64.sh
    1. Create an environment (here I use name np_veclib)
    $ conda create -n np_veclib python=3.9
$ conda activate np_veclib

    II. Install Numpy with BLAS interface specified as vecLib

    1. To compile numpy, first need to install cython and pybind11:
    $ conda install cython pybind11
    1. Compile numpy by (Thanks @Marijn's answer) - don't use conda install!
    $ pip install --no-binary :all: --no-use-pep517 numpy
    1. An alternative of 2. is to build from source
    $ git clone https://github.com/numpy/numpy
$ cd numpy
$ cp site.cfg.example site.cfg
$ nano site.cfg

    Edit the copied site.cfg: add the following lines:

    [accelerate]
libraries = Accelerate, vecLib

    Then build and install:

    $ NPY_LAPACK_ORDER=accelerate python setup.py build
$ python setup.py install
    1. After either 2 or 3, now test whether numpy is using vecLib:
    >>> import numpy
>>> numpy.show_config()

    Then, info like /System/Library/Frameworks/vecLib.framework/Headers should be printed.

    III. For further installing other packages using conda

    Make conda recognize packages installed by pip

    conda config --set pip_interop_enabled true

    This must be done, otherwise if e.g. conda install pandas, then numpy will be in The following packages will be installed list and installed again. But the new installed one is from conda-forge channel and is slow.

    Comparisons to other installations:

    1. Competitors:

    Except for the above optimal one, I also tried several other installations

    The above ABC options are directly installed from conda-forge channel. numpy.show_config() will show identical results. To see the difference, examine by conda list - e.g. openblas packages are installed in B. Note that mkl or blis is not supported on arm64.

    2. Benchmarks:

    Here I use two benchmarks:

    1. mysvd.py: My SVD decomposition
    import time
import numpy as np
np.random.seed(42)
a = np.random.uniform(size=(300, 300))
runtimes = 10

timecosts = []
for _ in range(runtimes):
    s_time = time.time()
    for i in range(100):
        a += 1
        np.linalg.svd(a)
    timecosts.append(time.time() - s_time)

print(f'mean of {runtimes} runs: {np.mean(timecosts):.5f}s')
    1. dario.py: A benchmark script by Dario Radečić at the post above.

    3. Results:

    +-------+-----------+------------+-------------+-----------+--------------------+----+----------+----------+
|  sec  | np_veclib | np_default | np_openblas | np_netlib | np_openblas_source | M1 | i9–9880H | i5-6360U |
+-------+-----------+------------+-------------+-----------+--------------------+----+----------+----------+
| mysvd |  1.02300  |   4.29386  |   4.13854   |  4.75812  |      12.57879      |  / |     /    |  2.39917 |
+-------+-----------+------------+-------------+-----------+--------------------+----+----------+----------+
| dario |     21    |     41     |      39     |    323    |         40         | 33 |    23    |    78    |
+-------+-----------+------------+-------------+-----------+--------------------+----+----------+----------+