## Hi Ed and Susan, are you using Intel MKL?

Bayesian analysis of blinking and bleaching

### Hi Ed and Susan, are you using Intel MKL?

Hi Ed and Susan,
are you using Intel MKL (Math Kernel Library)? ... if yes, it might be easy to get 3B to work on Phi card(s).

According to this blog
https://software.intel.com/en-us/articl ... sor-system
and more complete as a PDF file at
https://software.intel.com/sites/defaul ... Matlab.pdf

Intel now has "automatic optimization (AO)" that recognizes Phi cards.

Phi is looking even more fun - Intel plans to have a 3 Teraflop CPU version (72 cores, access to a lot of ram) in 2015:
http://vr-zone.com/articles/xeon-phi-kn ... 64112.html

I also note that NVidia's new TITAN Z GPU card is 8 Teraflops (and $3,000), so may be more useful to go for "3B-GPU(s)" instead of "3B-Phi(s)". Sincerely, George geomcnamara Posts: 4 Joined: Thu Jan 31, 2013 7:11 pm ### Re: Hi Ed and Susan, are you using Intel MKL? geomcnamara wrote:Hi Ed and Susan, are you using Intel MKL (Math Kernel Library)? ... if yes, it might be easy to get 3B to work on Phi card(s). Unfortunately from the point of view of this, we're not. The bulk of the CPU time is spent in some custom written routines that don't have inner loops that match any MKL functions. I also note that NVidia's new TITAN Z GPU card is 8 Teraflops (and$3,000), so may be more useful to go for "3B-GPU(s)" instead of "3B-Phi(s)".

What's interesting is the 3B code has a few serial bottlenecks in it, which are likely to limit the performance somewhat. I think porting 3B to a GPU could be a very interesting project, but it's hard to predict what the speedup will be in advance, unfortunately. I believe the Phi has a smaller number of faster processors which give it less overall performance but would make it less prone to being really slowed down by any serial bottlenecks.

Regards

-Ed
edrosten

Posts: 42
Joined: Wed Jul 18, 2012 2:54 pm

### Re: Hi Ed and Susan, are you using Intel MKL?

Hi Ed,

Bummer that Intel's MKL is apparently not going to help you with 3B for Phi ( https://software.intel.com/sites/defaul ... load_0.pdf ).

Did Hu et al 2013 OptNano http://www.optnano.com/content/2/1/7 get their speed up on a supercomputer by just using more cores?

Intel recently provided more details on Knights Landing, 3 Teraflop CPU with 16 Gb ram on (top of) the CPU
http://vr-zone.com/articles/intel-unvei ... 79686.html
which has a 30 minute web cast
http://www.brighttalk.com/webcast/10773/116329

The new information is similar to that in the late 2013 article I mentioned in my earlier post
http://vr-zone.com/articles/xeon-phi-kn ... 64112.html

FYI - the TACC Stampede I have free access held steady at #7, http://www.top500.org/site/1739 ... I am still trying to get help from our I.T. people on campus. I will be visiting TACC next week - they are hosting this year's Intel Xeon Phi User's Group Meeting
https://www.tacc.utexas.edu/news/events ... ly-meeting
and will encourage TACC'ers (TACCites?) to get 3B running themselves.

Sincerely,

George
geomcnamara

Posts: 4
Joined: Thu Jan 31, 2013 7:11 pm