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

Question

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:

• Why python run natively on M1 Max is greatly (~100%) slower than on my old MacBook Pro 2016 with Intel i5?
• On M1 Max, why there isn't significant speed difference between native run (by miniforge) and run via Rosetta (by anaconda) - which is supposed to be slower ~20%?
• On M1 Max and native run, why there isn't significant speed difference between conda installed Numpy and TensorFlow installed Numpy - which is supposed to be faster?
• On M1 Max, why run in PyCharm IDE is constantly slower ~20% than run from terminal, which doesn't happen on my old Intel Mac.

Evidence supporting my questions is as follows:

---

Here are the settings I've tried:

1. Python installed by

• Miniforge-arm64, so that python is natively run on M1 Max Chip. (Check from Activity Monitor, Kind of python process is Apple).
• Anaconda. Then python is run via Rosseta. (Check from Activity Monitor, Kind of python process is Intel).

2. Numpy installed by

• conda install numpy: numpy from original conda-forge channel, or pre-installed with anaconda.
• Apple-TensorFlow: with python installed by miniforge, I directly install tensorflow, and numpy will also be installed. It's said that, numpy installed in this way is optimized for Apple M1 and will be faster. Here is the installation commands:

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

3. Run from

• Terminal.
• PyCharm (Apple Silicon version).

---

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,

• run the same code on my old MacBook Pro 2016 with i5 chip - it costs 2.39917s.
• another post (but not in English) reports that run with M1 chip (not Pro or Max), miniforge+conda_installed_numpy is 2.53214s, and miniforge+apple_tensorflow_numpy is 1.00613s.
• you may also try on it your own.

Here is the CPU information details:

• My old i5:

$ 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

• My new M1 Max:

% 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 :)

Answer 1

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

3. 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

2. Compile numpy by (Thanks @Marijn's answer) - don't use conda install!

$ pip install --no-binary :all: --no-use-pep517 numpy

3. 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

4. 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

• A. np_default: conda create -n np_default python=3.9 numpy
• B. np_openblas: conda create -n np_openblas python=3.9 numpy blas=*=*openblas*
• C. np_netlib: conda create -n np_netlib python=3.9 numpy blas=*=*netlib*

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.

• D. np_openblas_source: First install openblas by brew install openblas. Then add [openblas] path /opt/homebrew/opt/openblas to site.cfg and build Numpy from source.
• M1 and i9–9880H in this post.
• My old i5-6360U 2cores on MacBook Pro 2016 13in.

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')

2. 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    |
+-------+-----------+------------+-------------+-----------+--------------------+----+----------+----------+

Answer 2

Possible Cause: Different BLAS Libraries

Since the benchmark is running linear algebra routines, what is likely being tested here are the BLAS implementations. A default Anaconda distribution for osx-64 platform is going to come with Intel's MKL implementation; the osx-arm64 platform only had the generic Netlib BLAS and the OpenBLAS implementation options when this question was first asked.

I get the following benchmark results (updated May 2023):

macOS Intel Core i9 (x86_64)

BLAS Implmentation	Mean Timing (s)
mkl	0.95932
blis	1.72059
openblas	2.17023
accelerate	2.56365
netlib	5.72782

macOS M1 4P/4E (arm64)

BLAS Implmentation	Mean Timing (s)
accelerate (macOS 13.3)	0.98718
accelerate (macOS 13.2)	1.03141
netlib	4.36523
openblas	10.33956

So, I suspect the old MBP had MKL installed, while the M1 system in OP is installing either Netlib or OpenBLAS. Apple Silicon users should identify which library runs fastest on their system. The consensus appears to be that for M1 and M2 systems, Apple's Accelerate library is the most performant.

It should be noted that Accelerate implementation was significantly updated in macOS 13.3 and this seems to provide a slight performance boost (~5%).

Hence, M1/M2 users should consider including the requirement:

'blas=*=accelerate'

when creating Conda environments.

---

Specifying BLAS Implementation

Here are specifically the different environments I tested:

## Note that `conda-forge` channel is prioritized on my system
## also, `mamba` is a faster version of `conda`

# MKL
mamba create -n np_mkl python=3.9 numpy 'blas=*=mkl'

# BLIS
mamba create -n np_blis python=3.9 numpy 'blas=*=blis'

# OpenBLAS
mamba create -n np_openblas python=3.9 numpy 'blas=*=openblas'

# Accelerate
mamba create -n np_accelerate python=3.9 numpy 'blas=*=accelerate'

# Netlib
mamba create -n np_netlib python=3.9 numpy 'blas=*=netlib'

and ran the benchmark script (so-np-bench.py) with

conda run -n np_mkl python so-np-bench.py

# etc.

---

Possible caveats with using Accelerate

Note that there has been incompatibility between SciPy and Accelerate due to the LAPACK implementation in Accelerate being old. However, it has been updated in macOS 13.3, which may resolve such issues. See this thread for details.

---

Emulation Mode (Rosetta)

Sometimes one may have to create an environment in emulation mode, e.g., need a package that doesn't have an osx-arm64 build yet. Here's the benchmarking (as of July 2023) using forced osx-64 on an M1:

macOS M1 4P/4E (arm64 with osx-64 subdir)

BLAS Implmentation	Mean Timing (s)
accelerate	1.71188
blis	2.41957
netlib	3.92932
openblas	5.98487
mkl	n/a (cannot emulate)

These runs use the same creation commands as osx-64 above, but with CONDA_SUBDIR=osx-64.

Important Note: The openblas runs use all cores, but to no apparent benefit. So I'd absolute recommend avoiding that for now. There might be some bug in how it is using OpenMP?

Takeaway: if you are on an Apple Silicon machine, use 'blas=*=accelerate' whether or not you are emulating. However, the native environment seems to be the fastest option.

Answer 3

With Miniforge3-MacOSX-arm64, and conda install -c conda-forge numpy "libblas=*=*accelerate", it works perfectly on my Macbook M1 Max.

• M1 Max with libblas accelerate : 1.024
• M1 Max without libblas accelerate : 2.672

Answer 4

Thank you for the tips. I followed the following in my freshly new MAC M1 MAX:

1. I installed the Minoforge3 (bash Miniforge3-MacOSX-arm64.sh)
2. Initialized a conda base environment (conda init) with Python 3.10
3. Installed numpy as: conda install numpy "libblas=*=*accelerate"

And then the suggested benchmarks from the link:

1. The script mention above mysvd.py runs in mean of 10 runs: 1.08088s

2. The script dario.py from https://gist.githubusercontent.com/daradecic/a2ac0a75d7e5f22c9aa07174dcbbe061/raw/a56ee217e6d3f949b1d1f719a7a134cef130cd9f/macs.py gives:

Dotted two 4096x4096 matrices in 0.28 s.
Dotted two vectors of length 524288 in 0.11 ms.
SVD of a 2048x1024 matrix in 0.44 s.
Cholesky decomposition of a 2048x2048 matrix in 0.07 s.
Eigendecomposition of a 2048x2048 matrix in 3.83 s.

TOTAL TIME = 19 seconds

Answer 5

Fastest way I found to do this is:

conda create -n np_accelerate -c conda-forge python=3.9 numpy 'blas=*=accelerate'

If you want to add conda forge for future use can use config append:

conda config --append channels conda-forge

conda create -n np_accelerate python=3.9 numpy 'blas=*=accelerate'

Other answers mention a few variants of this: blas=*=*accelerate* and libblas=*=*accelerate. These all work as well with the above command and result in same performance in my testing.