Seti is 10 times faster with Cuda, why not Leiden ?


advanced search

Message boards : Number crunching : Seti is 10 times faster with Cuda, why not Leiden ?

Reply to this thread
Subscribe to this thread
Sort
AuthorMessage
Zarck
private message
Joined: Feb 25, 2006
Posts: 6
ID: 451
Credit: 3,720
RAC: 38
Message 2715 - Posted 1 May 2011 13:36:34 UTC

Seti is 10 times faster with Cuda, why not Leiden ?

@+
*_*
____________

m.somers User profile image
Forum moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
private message
Joined: Nov 14, 2005
Posts: 662
ID: 1
Credit: 1,417,572
RAC: 2
Message 2716 - Posted 4 May 2011 8:29:57 UTC
Last modified: 4 May 2011 8:32:10 UTC

We do not have a CUDA / OpenCL app. There are several reasons for that;

1) GPUs are not always ieee compliant (see wiki http://en.wikipedia.org/wiki/GPGPU):

The implementations of floating point on Nvidia GPUs are mostly IEEE compliant; however, this is not true across all vendors.[1] This has implications for correctness which are considered important to some scientific applications. While 64-bit floating point values (double precision float) are commonly available on CPUs, these are not universally supported on GPUs; some GPU architectures sacrifice IEEE compliance while others lack double-precision altogether. There have been efforts to emulate double-precision floating point values on GPUs; however, the speed tradeoff negates any benefit to offloading the computation onto the GPU in the first place.[2]

This causes issues for us with homogeneous redundancy (the thing we use to validate calculations).

2) We do not possess the man-power to rewrite our codes using CUDA / OpenCL.

We are scientists and we want to do science and not so much programming. Some programming needs to be done, that is fine, but our focus is still on doing science. If we are able to do the science effectively without using GPUs, that is fine by us.

3) Using faster apps also means having (human) resources to analyze results faster.

With a shortage on good PhD students capable of doing science (understand the theoretical chemistry theories which are mostly quantum mechanics based) and capable of programming (have a keen interrest in computing and hardware and perhaps even in several languages; we find that we have 'scaling' issues on the human end and not so much on the computing side of things more and more.

4) We do not own any GPUs to test on and we certainly do not have the money to buy several. Also applying for grants to get that money is rather time-consuming and the success-rate is small.

All in all; it's economics; time v. shortage of budget v. shortage of highly scilled theoreticians or very smart programmers not being scared of learning or example something like quantum theory along the side.

m.




____________
M.F. Somers

Zarck
private message
Joined: Feb 25, 2006
Posts: 6
ID: 451
Credit: 3,720
RAC: 38
Message 2763 - Posted 24 Aug 2011 9:34:23 UTC - in response to Message ID 2716.

Amber Gpu Tools

Amber
____________

m.somers User profile image
Forum moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Avatar
private message
Joined: Nov 14, 2005
Posts: 662
ID: 1
Credit: 1,417,572
RAC: 2
Message 2764 - Posted 25 Aug 2011 9:53:05 UTC

Amber forcefield is not applicable to our science. Amber is not open source; deploying this on a grid or BOINC means trickery with licenses. Furthermore; one example of success does not prove generality ;-).

m.


____________
M.F. Somers

Dirk Broer
private message
Joined: Jul 12, 2009
Posts: 21
ID: 22421
Credit: 2,039,361
RAC: 1,269
Message 2768 - Posted 5 Oct 2011 12:56:29 UTC
Last modified: 5 Oct 2011 12:59:06 UTC

It is up to a project with a GPU application to decide which hardware requirements are needed in order to be able to participate in the GPU application. My Ati HD 3850 is suited for Collatz, DNETC and, by virtue of having double-precision capability, also for MilkyWay. It is however not suited for PrimeGrid because it does not support OpenCL.

64-bit floating point capability and IEEE compliancy might restrict the field even further, but there are people out there that have such cards, and some even might be able to write an OpenCL LeidenClassical application and test it for you.
____________

Reply to this thread

Message boards : Number crunching : Seti is 10 times faster with Cuda, why not Leiden ?



Return to Leiden Classical main page


Copyright © 2017 Leiden University - Leiden Institute of Chemistry - Theoretical Chemistry Department