MKL 11 in EPD

classic Classic list List threaded Threaded
6 messages Options
Reply | Threaded
Open this post in threaded view
|

MKL 11 in EPD

Vutshi
Hi everyone,

It is now almost a year since Intel released MKL 11. It is optimised for Ivy Bridge and the new Haswell CPUs.
Are there any plans to include it in EPD?
Reply | Threaded
Open this post in threaded view
|

Re: MKL 11 in EPD

Michael Aye
Bump!

On 2013-06-04 12:39:52 +0000, Vutshi said:

> Hi everyone,
>
> It is now almost a year since Intel released MKL 11. It is optimised for Ivy
> Bridge and the new Haswell CPUs.
> Are there any plans to include it in EPD?



_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL 11 in EPD

Vutshi
Dear Enthought,

recently, I did a comparison of MKL 10.3 from EPD and MKL 11 from Anaconda. On my laptop with core i7 (sandy bridge) the results are as follows

import numpy as np
N = 3000
a = np.random.rand(N, N)
c = np.random.rand(N, N)

%timeit np.dot(a,c)
MKL 10.3 gives 2.41 s
MKL 11 gives   1.22 s !!!

%timeit np.linalg.eig(a)
MKL 10.3 gives 35.3 s
MKL 11 gives    29.3 s

%timeit np.linalg.eigh(a + a.T)
MKL 10.3 gives 8.45 s
MKL 11 gives    6.28 s

%timeit np.linalg.inv(a)
MKL 10.3 gives 3.77 s
MKL 11 gives    2.27 s

%timeit np.linalg.pinv(a)

MKL 10.3 gives 25.7 s
MKL 11 gives    19.5 s

This is quite a good improvement I would say and newer CPUs should gain even more. I'd like to see it in EPD!
Reply | Threaded
Open this post in threaded view
|

Re: MKL 11 in EPD

Frédéric Bastien
Hi,

There have been optimization in numpy too from memory. Can you tell
witch version of numpy was installed on each just to make sure this
isn't the reason.

For example, the speed up from numpy.dot() could be due to that.

Fred

p.s. I'm not working with Enthought, but I follow the numpy
development for our project Theano.

On Wed, Sep 25, 2013 at 7:49 AM, Vutshi <[hidden email]> wrote:

> Dear Enthought,
>
> recently, I did a comparison of MKL 10.3 from EPD and MKL 11 from Anaconda.
> On my laptop with core i7 (sandy bridge) the results are as follows
>
> import numpy as np
> N = 3000
> a = np.random.rand(N, N)
> c = np.random.rand(N, N)
>
> %timeit np.dot(a,c)
> MKL 10.3 gives 2.41 s
> MKL 11 gives   1.22 s !!!
>
> %timeit np.linalg.eig(a)
> MKL 10.3 gives 35.3 s
> MKL 11 gives    29.3 s
>
> %timeit np.linalg.eigh(a + a.T)
> MKL 10.3 gives 8.45 s
> MKL 11 gives    6.28 s
>
> %timeit np.linalg.inv(a)
> MKL 10.3 gives 3.77 s
> MKL 11 gives    2.27 s
>
> %timeit np.linalg.pinv(a)
>
> MKL 10.3 gives 25.7 s
> MKL 11 gives    19.5 s
>
> This is quite a good improvement I would say and newer CPUs should gain even
> more. I'd like to see it in EPD!
>
>
>
> --
> View this message in context: http://enthought-dev.117412.n3.nabble.com/MKL-11-in-EPD-tp4026502p4026679.html
> Sent from the Enthought Dev mailing list archive at Nabble.com.
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: MKL 11 in EPD

Vutshi
The Numpy version is 1.7.1 for both EPD and Anaconda.
Reply | Threaded
Open this post in threaded view
|

Re: MKL 11 in EPD

Vutshi
Apparently, there is MKL 11.1 already available.
http://software.intel.com/en-us/forums/topic/472068