MKL 11 Versus MKL 10 Performance

MKL 11 Versus MKL 10 Performance

Аватар пользователя Jason S.
Jason S.
ср, 20/02/2013 - 15:12

I have a benchmark application that mainly does two types of things:

* FFTs

* Eigenvector decompositions

In both cases, the new MKL11 benchmarks much slower than MKL10 on a I7 IVY Bridge, Sandy Bridge and other CPUS.

For the Eigenvector code, it is about 1/2 the performance in version 11.  For the FFTs the performance is degraded by about 30%.

Is there a reason why MKL 11 performs worse?  Is there something I need to do to at least get equivalent performance to MKL 10?

Thanks for your assistance in advance.

RSS Вверх
6 сообщений / 0 новое
Последнее сообщение
Пожалуйста, обратитесь к странице Уведомление об оптимизации для более подробной информации относительно производительности и оптимизации в программных продуктах компании Intel.
Аватар пользователя Chao Y (Intel)
Chao Y (Intel)
ср, 20/02/2013 - 18:32

Hello,

could you please a sample code show this performance problem?  Also let us know 1)how code is linked 2) Is it a 32 bit code or Intel64 code 3) Windows/Linux?

Regards,
Chao

Вверх
Аватар пользователя Jason S.
Jason S.
чт, 21/02/2013 - 09:51

I have attached a complete VS2008 project which demonstrates the problem.  Press the "EigenValues" button and  "Run Benchmark" buttons to see the performance.  As a correction the FFT times are improved in version 11 BUT the eigenvalue calculation times are substantially worse under version 11.  Here is a summary of measured benchmarks:

EIGENBenchmark:

  I7-920 32Bit MKL10: 1.3 Seconds; I7-920 32Bit MKL11: 2.53 Seconds; I7-920 64Bit MKL10: 1.09 Seconds; I7-920 64Bit MKL11: 1.65 Seconds

 I7-3770 32Bit MKL10: .757 Seconds; I7-3770 32Bit MKL11: 1.4 Seconds; I7-3770 64Bit MKL10: .63 Seconds, I7-3770 64Bit MKL1: .98 Seconds

In order to compile, you copy the noted (in .txt file) .lib files into the folders for libraries by the way.

 

Вверх
Аватар пользователя Jason S.
Jason S.
чт, 21/02/2013 - 09:53

So the mystery is: why is MKL11 substantially worse in performance for Eigenvector decompositions compared to MKL 10.

Regards

Вверх
Аватар пользователя Jason S.
Jason S.
чт, 21/02/2013 - 09:56

I sent the project and source code as an attachment, but I dont see it in the chat.  Is there some magic I need to do to get the attachment uploaded?  Should I email it somewhere?

Вверх
Аватар пользователя Sergey Kostrov
Sergey Kostrov
вс, 24/02/2013 - 16:34

>>...I sent the project and source code as an attachment, but I dont see it in the chat...

Why wouldn't you add a zip-file with the test project to a new post in the current thread?

In that case everybody will be able to test, if interested. As soon as you add the zip-file press 'Start upload' button and as soon as download completed press 'Submit' button.

Note: There is a little problem. A developer should have MKL v10 and MKL v11 installed on its computer. I don't have MKL v10, for example.

Вверх

Зарегистрируйтесь, чтобы оставить комментарий.