This is an archival dump of old wiki content --- see for current material

Introduction : source and binaries

Building from sources

To build Numpy/Scipy from source, get the source package, unpack it, and:

python install --user   # installs to your home directory -- requires Python >= 2.6


python build
sudo python install --prefix=/usr/local   # installs to /usr/local

Before building, you will also need to install packages that Numpy and Scipy depend on

Typically, you will want to install all of the above from packages supplied by your Linux distribution, as building them yourself is complicated. If you need to use specific BLAS/LAPACK libraries, you can do

export BLAS=/path/to/
export LAPACK=/path/to/
export ATLAS=/path/to/
python ............

If you don't want to any LAPACK, just do "export LAPACK=".

You will find below additional installation instructions and advice for many major Linux distributions.

Table of Contents

Fedora Core 8, openSUSE 10.2, RHEL/Centos 5

I (DavidCournapeau) have packaged the last released of numpy, scipy as well as lapack and blas dependencies for Fedora Core 8, opensuse 10.2 and Centos/RHEL 5 and a few others thanks to the opensuse build service. I strongly advise you to use those packages instead of the "official" ones, which are often unusable. The repository is there:

Basic install

To use this repository with yum, simply pick up your arch/distribution from , and take the corresponding .repo file. Put this .repo file into /etc/yum.repo.d/, and then install numpy/scipy with yum:

yum install python-numpy python-scipy

I also packaged timers and testers for blas and lapack, which can be useful if you intend to compile special optimized versions of BLAS/LAPACK (eg GOTO or ATLAS). You can also find the package lapack3-pic, which can be used to build a complete LAPACK with ATLAS: it is a static version, but as it is built with the -fPIC compiler flag, it can be used to build python extensions; this is particularly useful for x86_64 arch.

Building Atlas

ATLAS is a BLAS/LAPACK implementation which tuned itself on the machine to provide ideal performances, and often match vendor specific implementations. Unfortunately, building ATLAS is not easy. But, it is getting easier all the time.

Building Atlas by Hand

These instructions show how to build ATLAS (and LAPACK) from their official distributions.

First, download and unpackage the LAPACK distribution from netlib (you need these to build a complete version of LAPACK).

tar zxvf lapack-3.1.1.tgz
cd lapack-3.1.1

There are several files in the INSTALL directory of the lapack distribution. Copy one of those files to the main directory. For example:


Now, you must edit the file to ensure that the OPTS and NOOPT lines both contain the flag for compiling position-independent code on your platform (e.g. with gcc/gfortran it is -fPIC). For example:

OPTS = -O2 -fPIC

(Note: Make sure that if you build with gfortran that g77 is not installed on your system (or at least is not in your PATH when numpy is being built) as you need to link with the same compiler that you built lapack with when numpy builds. It will try and find g77 first which will lead to linking errors if you have built lapack with gfortran). Then change to the SRC directory and run make

cd SRC

This will create an lapack_<XXXX>.a file in the head lapack directory. You will need the location of this file to configure atlas.

Now, download the latest release of ATLAS (these instructions worked on 3.7.37). See, for example, Unpackage the result, change to the directory created, and create a directory to contain the resulting build. This directory should be named appropriate for the platform (you can build for multiple platforms from the same SOURCE tree --- perhaps the source is on a network drive and builds are taking place for multiple platforms).

tar jxvf atlas3.7.37.tar.bz2
mkdir ATLAS_<my_platform_type>

cd ATLAS_<my_platform_type>
../configure -Fa alg -fPIC --with-netlib-lapack=/path/to/lapack/lapack_<XXXX>.a

Your atlas libraries should now be in the lib subdirectory of the current directory. You should copy them to some-place that you can tell site.cfg about so that numpy and scipy can pick them up. If you want to create shared libraries, then you can do that by

cd lib
make shared     # for sequential libraries
make ptshared   # for threaded libraries

after changing to the lib directory where the .a files are already located.

Building Atlas with atlas-XXX.src.rpm from Ashigabou Repository

ashigabou repository does not provide binary versions, but provides all the tools to make the building process of ATLAS almost painless: it will build a complete LAPACK, build it with the right fortran compiler to avoid ABI issues (eg _gfortran_string_write, etc...), and with the right compiler flags such as it is usable to build numpy and scipy (with the -fPIC option).

First, download the source rpm included in the ashigabou repository (the file atlas-version.src.rpm), and install from the ashigabou repository the package lapack3-pic (the rpm will refuse to build without it). Then, use the following:

rpm -ivh atlas-version.src.rpm

This will NOT install atlas, just uncompress all the necessary files for building the rpm in /usr/src/packages. Before building atlas, you must disable dynamic change of CPU frequency (used to decrease battery consumption):

cpufreq-selector -g performance

If this fails telling you no cpufreq support, this is fine. Now, to build the rpm, go into the directory /usr/src/packages/SPEC, and execute

rpmbuild -ba atlas.spec

This will build the rpm: this can take a long time, even on a powerful machine. What matters is whether atlas has arch defaults for your machine: if not, it can take several hours (it takes 2 hours and a half on a P4 @3.2 Ghz, but takes ~10 minutes on my macbook under linux). If successfull, you will get an installable rpm in /usr/src/packages/RPMS/ARCH (where ARCH can be x86_64 or i586 or something else depending on the distribution and your arch).

The rpm contains two (shared) libraries: and, installed in /usr/lib/atlas/sse2. They are meant to be drop-out for the standard BLAS and LAPACK (the ones in refblas3 and lapack3). To use the atlas libraries, once you installed numpy and scipy, you should tell the OS to use atlas instead of default libraries by using LD_LIBRARY_PATH. That is, normally, you can use numpy by :

python -c "import numpy as N; a=N.random.randn(1000, 1000);, a)"

To use atlas, you do:

LD_LIBRARY_PATH=/usr/lib/atlas/sse2 python -c "import numpy as N; a=N.random.randn(1000, 1000);, a)"

If everything is working correctly, you will see that the above script runs much faster with atlas than without (I see a ten fold speed increase on my machine).

Mandriva 2007.1

Binary packages for NumPy and SciPy are available via the contrib urpmi repository:

   urpmi python-scipy


Gentoo includes an ebuild. Type:

   sudo emerge scipy

Debian / Ubuntu

Debian and Ubuntu ship with Numpy and Scipy -- to install their binary packages, use

sudo apt-get install python-numpy python-scipy

Note (esp. Ubuntu versions prior to Maverick): Do not install versions 3.6.0-* of libatlas-sse2 or libatlas-sse packages -- they contained severe known bugs.

Andrew Straw's unofficial repository for Ubuntu / Debian

Andrew Straw has an unofficial repository for NumPy .deb packages. These were built with stdeb. The binaries are for Ubuntu Dapper (6.06 LTS).

Binary packages Ubuntu Dapper (6.06), (i386 and amd64 architectures)

To use the binary package in Ubuntu Dapper, add the following line to your /etc/apt/sources.list:

deb dapper/

Then type

sudo apt-get install python-numpy

You can verify ATLAS support by running the command ldd /usr/lib/python2.4/site-packages/numpy/linalg/, which should result in output like the following: => /usr/lib/atlas/ (0x00002aaaaabcf000) => /usr/lib/atlas/ (0x00002aaaab435000) => /usr/lib/ (0x00002aaaabd15000) => /lib/ (0x00002aaaabe44000) => /lib/ (0x00002aaaabfca000) => /lib/ (0x00002aaaac0d7000)
        /lib64/ (0x0000555555554000)

Source packages for any Debian-based distribution

The following may (or may not) work on any Debian-based distribution:

Add the following line to your /etc/apt/sources.list:

deb-src dapper/

To download and build, type:

sudo apt-get build-dep python-numpy
sudo apt-get -b source python-numpy

GPG Verification using Andrew Straw's repository

When you start using this repository, you might get warning messages like this:

The following signatures couldn't be verified because 
the public key is not available.

Or you will be asked questions like this over and over:

WARNING: The following packages cannot be authenticated!
Install these packages without verification [y/N]?

Install the package astraw-keyring to eliminate these messages. This installs Andrew's archive signing key to your apt through the apt-key add command.

Debian sarge notes

If you install NumPy or SciPy ontop of a debian sarge installation for a CPU with SSE2, there is a bug in libc6 2.3.2 affecting floating point operations (fixed in version 2.3.3). Due to this bug, the numpy and scipy tests crach with a SIGFPE. Since there is now patch available, in order to fix this the libc6 sources need to be downloaded, fixed, and rebuilt. See Andrew Straw's instructions for more information.

Ubuntu notes

If you choose not to use Andrew Straw's repository (which includes numpy built with ATLAS support), here are some further notes to build numpy and scipy from sources on your computer.

First, you need to install several libraries/tools (you need to enable universe repository for some of those packages):

sudo apt-get install gcc g77 python-dev atlas3-base-dev

To use optimized lapack and blas, you should also install the atlas corresponding to your achitecture: atlas3-sse2-dev if you have a CPU with SSE2 capabilities, atlas3-sse-dev if you have a CPU with SSE capabilities only, etc... If you have a recent x86 (eg intel or AMD cpu), it should support SSE2. To check whether your CPU supports sse, sse2, etc.. you can check using the following command:

cat /proc/cpuinfo | grep flags

and check whether sse, sse2, etc... appear on it.

Then, you can build numpy with the following, inside the numpy source directory:

python build

Then, to install it system-wide (requires root privileges):

python install

To install it in another directory, you need to use the prefix option. For example, I like to install local softwares in my $HOME/local, so I do the following:

python install --prefix=$HOME/local

Note that if you do not install numpy system wide, you need to tell python to look for the directory where you installed numpy. For example, if you use $HOME/local as the former example, then you should add $HOME/local/lib/python2.4/site-packages in your PYTHONPATH:

PYTHONPATH=$HOME/local/lib/python2.4/site-packages python

(change python2.4 to python2.5 if you are using python2.5, obviously).


(This section reflects the situation of July 2009. If you have newer of more accurate information, feel free to modify this section.)

OpenSUSE does not contain Numpy, Scipy or Matplotlib in the standard installation. Instead those packages are provided by additional repositories, that seem to be run by volunteers. However Novell provides webspace for some of those repositories. Packages usually exist only for a few current SUSE versions.

The following repositories are currently the best to obtain Numpy, Scipy and Matplotlib. They can be added to the package manager (YaST) with the Installation Source dialog. The packages will then appear in the Software Management dialog.

Alternatively the *.rpm files can be downloaded and installed manually (for example 'rpm -U <filename>' or with 'kpackage').

Alternatively one can search for packages in repositories hosted by Novell here:

One can also search for packages in the very big Packman repository:

Users of older versions of SUSE/openSuse can install Sage, a big collection of Mathematics related software. It was recently (Jul. 2009) reported that compiling and installing Sage from sources worked flawlessly, on SUSE Linux 10.2:


ATLAS is a replacement for BLAS and parts of LAPACK, that is much faster. It must be built from sources, because it optimizes itself for the computer's processor. The build process will run for ten minutes to several hours.

There is currently no comfortable way to use ATLAS on openSuse.

The build instructions for ATLAS on this page work, but unfortunately the Numpy and Scipy packages don't work with ATLAS. One could build Numpy and Scipy from sources though, and a relatively painless way to do this is the Sage package. (If you know a comfortable way to make ATLAS work on openSuse, please put it here into the Wiki.)

David Cournapeau has a repository devoted to ATLAS, but he has not added packages for recent SUSE versions.

Broken BLAS

SUSE (and Red Hat) regularly shipped versions of the BLAS library where some functions were missing. This bug has finally been fixed in March 2007. This means SUSE 10.2 and prior come with a broken BLAS, in later versions SUSE's original BLAS should work. Unfortunately the repositories mentioned here do no longer contain corrected/complete packages of BLAS and LAPACK for the affected versions (SUSE 10.2 and older).

The bug's cause was as follows: The BLAS rpm is created from Netlib's LAPACK package and not from the BLAS package. Until March 2007 however the LAPACK library did only contain a subset of the functions that were in BLAS. Finally someone begged the LAPACK developers to include the whole BLAS library in the LAPACK package, and they did.

For details see: and

Building everything from source with gfortran on Ubuntu (Nov 2010)

These are instructions for building everything from source on a 64 bit Ubuntu system (Maverick: 10.10) on a multicore processor using the latest versions as of November 2010. Everything is installed in a user directory structure in $HOME/local (/home/sam/local in my case). Administrator priviliges are required only in the beginning to disable CPU throttling while building ATLAS.

Install required packages

sudo apt-get install build-essential python-dev swig gfortran python-nose

Step 1: Disable CPU Throttling

ATLAS' timing algorithm require CPU throttling to be disabled. This disables it on the 0th core:

sudo cpufreq-selector -g performance

Then disable it on each additional core. For a quad core processor, these commands will be required:

sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor 
sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor 
sudo cp /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor /sys/devices/system/cpu/cpu3/cpufreq/scaling_governor 

Step 2: Build ATLAS(3.9.32) with complete Lapack(3.2.2)

Download lapack.tgz from and atlas3.9.32.tar.bz2. Extract the atlas archive into a directory named ATLAS and from within it, issue these commands:

mkdir BUILD
../configure -b 64 -Fa alg -fPIC --with-netlib-lapack-tarfile=<path to lapack.tgz> --prefix=/home/sam/local 
cd lib
make shared
make ptshared
cd ..
make install

Note that make ptshared might not work on a single core machine. Note also that the first "make" command above will take several hours to run, as ATLAS optimizes various performance parameters.

Step 3: Build UMFPACK (5.5.0) with AMD (2.2.1), UFConfig(3.5.0) and CHOLMOD (1.7.1)

NOTE: Dowloading and building the entire SuiteSparse all at once is easier than the following steps. SuiteSparse Version 4.0.2 is confirmed to have worked on Xubuntu 12.10 with the latest ATLAS, LAPACK, NumPy and SciPy as of 2012-10-18. --MartinSpacek


Dowload all four packages and extract them in the same directory. Edit UFconfig/ to read:

CC = gcc
CFLAGS = -O3 -fexceptions -m64 -fPIC

F77 = gfortran
F77FLAGS = -O -m64 -fPIC

INSTALL_LIB = /home/sam/local/lib
INSTALL_INCLUDE = /home/sam/local/include



Then issue the following commands

make library
make install
cd ../AMD
make install
cd ../UFconfig
cp UFconfig.h /home/sam/local/include

Step 4: Build numpy(1.5.0)

Untar the archive, copy site.cfg.example to site.cfg and edit it:

library_dirs = /home/sam/local/lib
include_dirs = /home/sam/local/include

In the same file, uncomment these lines:

libraries = ptf77blas, ptcblas, atlas

libraries = lapack, ptf77blas, ptcblas, atlas

amd_libs = amd

umfpack_libs = umfpack

For a single core machine, uncomment these lines:

libraries = f77blas, cblas, atlas
libraries = lapack, f77blas, cblas, atlas

Then use the standard installation technique

python build
python install --prefix=/home/sam/local

Step 5: Build scipy(0.8.0)

Make sure that $HOME/local/bin is in $PATH (for f2py) and $PYTHONPATH contains $HOME/local/lib/python-2.6/site-packages (for numpy)

Do a standard install

python build
python install --prefix=/home/sam/local

Building everything from source with gfortran on Ubuntu (Feb 2008)

This is how I built everything from source on a 64 bit Ubuntu system with latest versions as of February 2008. It took me some time to work out the issues so I thought I'd put the details here. I believe it should work the same on 32 bit systems (leaving out 64 bit related options).

Install required packages

sudo apt-get install build-essential python-dev swig gfortran

Install nose (easy_install nose). Do not install python-nose, it is an earlier version that doesn't work with scipy. Also make sure g77 is not installed. Distutils will not use gfortran if g77 is installed.

sudo apt-get remove python-nose
sudo apt-get remove g77
sudo apt-get install python-setuptools
sudo easy_install nose

Build lapack (3.1.1)

As described above, copy, add -fPIC flags (and -m64 if building 64 bit) to OPTS and NOOPT. Run make in SRC directory.

Build ATLAS (3.8.0)

As described above untar, create a directory for your build in ATLAS and run configure (add option '-b 64' for 64 bit).

sudo cpufreq-selector -g performance
../configure -b 64 -Fa alg -fPIC --with-netlib-lapack=/path/to/lapack/lapack_<XXXX>.a

Copy the libraries to a lib directory (/usr/local/lib or ~/scipy_build/lib for example). I found it's easier to copy all needed libraries and files to a common directory.

Build UMFPACK (5.2.0)

Get the latest versions of AMD, UFconfig and UMFPACK and untar them into a directory.

UFconfig/ should contain:

CC = gcc
CFLAGS = -O3 -fexceptions -m64 -fPIC
F77 = gfortran
F77FLAGS = -O -m64 -fPIC

BLAS = -L/usr/lib/gcc/x86_64-linux-gnu/4.2.1 -L/home/robince/scipy_build/lib -llapack -lf77blas -lcblas -latlas -lgfortran
LAPACK = -L/usr/lib/gcc/x86_64-linux-gnu/4.2.1 -L/home/robince/scipy_build/lib -llapack -lf77blas -lcblas -latlas -lgfortran

On a 32 bit system, remove the -m64 flags and change the first -L option to -L/usr/lib/gcc/i486-linux-gnu/4.2.1.

Run 'make' in UMFPACK directory. Copy resulting libraries and include files.

cp AMD/Lib/libamd.a ~/scipy_build/lib
cp UMFPACK/Lib/libumfpack.a ~/scipy_build/lib
cp AMD/Include/amd.h ~/scipy_build/lib/include
cp UFconfig/UFconfig.h ~/scipy_build/lib/include
cp UMFPACK/Include/*.h ~/scipy_build/lib/include

Copy libgfortran into scipy library directory (doesn't seem to work if it doesn't find the umfpack_libs together).

cp /usr/lib/gcc/x86_64-linux-gnu/4.2/libgfortran.* ~/scipy_build/lib/

Build FFTW (3.1.2)

SciPy Versions >= 0.7 and Numpy >= 1.2

Because of license, configuration, and maintenance issues support for FFTW was removed in versions of SciPy >= 0.7 and NumPy >= 1.2. Instead now uses a built-in version of fftpack.

There are a couple ways to take advantage of the speed of FFTW if necessary for your analysis.

  1. Downgrade to a Numpy/Scipy version that includes support.
  2. Install or create your own wrapper of FFTW. See as an un-endorsed example.

SciPy Versions < 0.7 and Numpy < 1.2

After untarring, run configure. I ran configure first and extracted the suggested FLAGS from the Makefile, then added -fPIC and -m64. (Not sure if this is necessary)

./configure --enable-sse2 --enable-threads --with-combined-threads CFLAGS="-O3 -fomit-frame-pointer -fstrict-aliasing -ffast-math -pthread -fPIC -m64" FFLAGS="-g -O2 -fPIC -m64" CXXFLAGS="-g -O2 -fPIC -m64"
sudo make install

Build Numpy and Scipy

Set the following entries in site.cfg (this will also work with fftw if it has been compiled and installed in the default location (/usr/local):

library_dirs = /usr/local/lib:/home/robince/scipy_build/lib
include_dirs = /usr/local/include:/home/robince/scipy_build/lib/include

atlas_libs = lapack, f77blas, cblas, atlas

amd_libs = amd

umfpack_libs = umfpack, gfortran

libraries = fftw3

Build Numpy and Scipy.

python build
sudo python install

Any distribution with Intel C compiler and MKL

The Intel C compiler and Intel MKL are free for personal non-commercial use on Linux.

Intel MKL 11.0 (updated Dec 2012)

Add the following lines to site.cfg in your top level NumPy directory to use Intel® MKL for Intel® 64 (or earlier known as em64t) architecture, considering the default installation path of Intel® MKL which is bundled with Intel® Composer XE SP1 version on Linux:

library_dirs = /opt/intel/composer_xe_2013/mkl/lib/intel64
include_dirs = /opt/intel/composer_xe_2013/mkl/include
mkl_libs = mkl_intel_lp64,mkl_intel_thread,mkl_core

If you are building NumPy for 32 bit, please add as the following

library_dirs = /opt/intel/composer_xe_2013/mkl/lib/ia32
include_dirs = /opt/intel/composer_xe_2013/mkl/include
mkl_libs = mkl_intel,mkl_intel_thread,mkl_core

Instead of the layered linking approach for the Intel® MKL as shown above, you may also use the dynamic interface lib mkl_rt.lib. So, for both the ia32 and intel64 architecture make the change as below

mkl_libs = mkl_rt

Modify cc_exe in numpy/numpy/distutils/ to be something like:

cc_exe = 'icc -O2 -g -openmp -avx'

Here we use, default optimizations (-O2), OpenMP threading (-openmp) and Intel® AVX optimizations for Intel® Xeon E5 or E3 Series which are based on Intel® SandyBridge Architecture (-avx). Run icc --help for more information on processor-specific options.

Compile and install NumPy with the Intel compiler (on 64-bit platforms replace "intel" with "intelem"):

python config --compiler=intel build_clib --compiler=intel build_ext --compiler=intel install

Compile and install SciPy with the Intel compilers (on 64-bit platforms replace "intel" with "intelem"):

python config --compiler=intel --fcompiler=intel build_clib --compiler=intel --fcompiler=intel build_ext --compiler=intel --fcompiler=intel install

You'll have to set LD_LIBRARY_PATH to Intel® MKL libraries (exact values will depend on your architecture, compiler and library versions) and OpenMP library for NumPy to work. If you build NumPy for Intel® 64 platforms:

$export LD_LIBRARY_PATH=/opt/intel/composer_xe_2013/mkl/lib/intel64: /opt/intel/composer_xe_2013/compiler/lib/intel64:$LD_LIBRARY_PATH

If you build NumPy for ia32 bit platforms:

$export LD_LIBRARY_PATH=/opt/intel/composer_xe_2013/mkl/lib/ia32: /opt/intel/composer_xe_2013/compiler/lib/ia32:$LD_LIBRARY_PATH

Intel MKL 10.0

The above instructions must be slightly modified to install NumPy 1.6 with Intel MKL 10.0 on a 64-bit Red Hat 4 system. If threading is desired, set

mkl_libs = mkl_intel_lp64, mkl_intel_thread, mkl_core, guide

Ensure that the -openmp compile flag is passed to the Intel Fortran Compiler (NOT the C compiler).

If threading is not desired, set

mkl_libs = mkl_intel_lp64, mkl_core.

Older information

It is possible that LD_LIBRARY_PATH causes a problem, if you have installed MKL and Composer XE in other directories than the standard ones. The only solution I've found that always works is to build Python, NumPy and SciPy inside an environment where you've set the LD_RUN_PATH variable, e.g:

export LD_RUN_PATH=~/opt/lib:~/intel/composer_xe_2013/compiler/lib:~/intel/composer_xe_2013/mkl/lib/ia32

Configure Python with --prefix=$HOME/opt, make, make install, add $HOME/opt/bin to the front of your PATH and then build NumPy and SciPy with the site.cfg as above in their top level directories (check the config step's output carefully to make sure it selects MKL). Built like this, you shouldn't have to set any LD_LIBRARY_PATH for NumPy and SciPy to work. Run the test suites to verify this.

Any Linux distro: self-contained local installation with Sage.

All you need is some basic tools like gcc (no fortran).

Follow the instructions here to build sage from source:

All you have to do is unpack the tar and type make. This takes about 3 hours.

This will install sage in its own directory. python (and ipython) can be found in SAGEROOT/local/bin

If you don't want to have to type in absolute paths, you can set the environment variables to point to your sage executables. To do this, run sage with the -sh option. My .profile contains the line

~/Sage/sage -sh


SciPy: Installing_SciPy/Linux (last edited 2015-10-24 17:48:24 by anonymous)