Hi everyone, I am sure this has been asked before... but where is the GPU support for ClimatePrediction.net?

As things are are now, the app is pretty stable so couldn't the Met just forgoe the effort into updates that would normally come and instead focus on getting support for ATI and nVidia cards for a short while?

Look at some of the figures here, BOINC forums suggest than credits work universally and so each project doesnot award points through credits which have differing scales for the amount of work done.

CPU Projects:
Einstiein - 1282.48
SETI - 4245.50
Cosomology - 2520.00

GPU Projects:
Collatz - 208513.08
DNETC - 142324.00
Milkyway - 266558.66

So, the throughput of even old GPU's being far quicker than CPU's like Intel's new i7 is not extactly breaking news. It is open for debate how credits are issued across different projects as I have not heard anything on the subject from the project admins.

It is also known that CPU's although slower are more accurate, however the sheer amount of work the GPU's get through would render that argument invalid. By the time my 4.0GHz, 8 thread i7 completed 1 work unit... the 2 HD5870's could have completed work units going into double figures.

Can the forum admins shed any light into the Met Offices plans for future support of GPU's? There would be alot of interest from the community and it would be nice to see these work units with an estimated finish time somewhere around the two hour mark.

:D

There are several threads about GPUs on this board.

And while the source programs ARE from the Met office, the sub programs which form the individual types of model are created/modified by the 2 software engineers who work on this project.
As this project is part of the research of Oxford University's Atmospheric, Oceanic & Planetary Physics, Department of Physics, the people involved will no doubt be in touch with the Met office.

But as to the plans of the Met Office, you'll need to ask them that. Especially in light of the massive spending cuts about to be imposed on the UK.

This project uses very large data bits for accuracy, and current GPUs and their programs just aren't up to it.
So don't hold your breath waiting for that sort of work here.

---

Backups: Here

I'm afraid that the CPDN researchers don't care one little bit about speed of crunching. All they care about is whether the models have been processed stably and produce valid results.

When the long BBC models were optimised because some of them were generating processing errors, they were slowed down by 20%. The current FAMOUS models were also slowed down to reduce the incidence of processing errors.

So really the only thing that matters is the validity of the results. I've heard that the way that climate models process the data doesn't lend itself to the way GPUs work. There are several centres in the world that process climate models and as far as I know none have adopted GPU processing. Not yet, anyway. The situation may of course change.

To my knowledge the only Boinc project that has implemented the New Credits system is Seti. When CPDN joined Boinc its credit awards were calibrated against what Seti was then awarding ie the number of credits per crunching hour on the same sort of computer, CPU-only. Since then some model types have been awarded credits at a slightly higher level to reward members for the large upload and download file transfers. And two years ago all model types had their credits increased by 5%, largely because of the upward credit trend right across Boinc projects.

As far as I know there are no plans at the moment to move CPDN to New Boinc Credits. I think CPDN's current Boinc version allows the comparison of New and Old (=current) for our crunching. I don't know whether any such comparison has been made yet, or the data for it even collected.

---

Cpdn news

Scientists at various atmospheric modeling centers are looking at using GPUs to process weather/climate models, or at least parts of the model calculation. Numerous difficulties exist in porting Fortran code to CUDA, in memory optimization, in ensuring the validity of the results. No doubt these will be worked out in time and GPUs will be used for at least calculating parts of climate models. However, a lot of work is yet to be done, and decisions about where scarce resources should be used have to be made. Where the Met Office is at with regard to this is a mystery to me. Even if the Met Office were to hand over usable code tomorrow to the cpdn programmers, getting it to work cross platform with various model types, validating the results, ensuring reasonable stability, etc. would take a long time. Then there's the whole "one or two cpdn programmers" and how their time is divided. Right now it is working on multiple model types and experiments for the numerous scientists with ideas on what experiments would be useful.

A couple articles about this, with some detail on some of the difficulties involved.

http://www.ciara.fiu.edu/publications/EScience2010.pdf
http://www.mmm.ucar.edu/wrf/WG2/michalakes_lspp.pdf