Well, Nvidia released it's latest line of GPUs yesterday and they have included a feature they are calling CUDA. Here's a brief explanation:

"CUDA (Compute Unified Device Architecture) takes the programmability of general-purpose computing, adds some dedicated caches, and places it in the context of the graphics chip. Run as a graphics chip, the 128 1.35-GHz thread processors process graphics data. When switched to run as a GPU computing element, however, the same elements rapidly calculate computational algorithms."

I think this is a great opportunity for Seti/BOINC programmers to step up and create something that will allow the GPU to help process work units. I've ordered my GTX and am ready to crunch. Just worried about my power bill.

---

While this topic has been massaged many times before (see discussion about Folding@home running on ATI graphics cards), I believe this development is finally in the right direction. Nokia is going to make SDK available "to those whom it should go to". I hope that includes Seti@home!

Here is an excellent article which elaborates on CUDA implementation.

---

With this new GPU it might be possible to code everything for the GPU only (or even several "cpus" if I read the specs correctly).

With current GPUs, the best that can easily be done is to compute the FFT in the GPU. There is already source for that (open source). This begins to get a bit tricky on systems with 1 Graphics board and 4 cpus (cores, threads, whatever).

In this situation the code would have to split the time of the GPU up among the 4 running seti programs, periodically crunching the FFTs and delivering it to each.

There are some major memory bandwidth problems getting the data up into to the GPU memory and back down again.

Once I break FFT out into its own class with versions for each current FFT method (Intel, FFTW, Ooura) it should be easier to add GPU FFT as a test situation.

With this new GPU it might be possible to code everything for the GPU only (or even several "cpus" if I read the specs correctly).

With current GPUs, the best that can easily be done is to compute the FFT in the GPU. There is already source for that (open source). This begins to get a bit tricky on systems with 1 Graphics board and 4 cpus (cores, threads, whatever).

In this situation the code would have to split the time of the GPU up among the 4 running seti programs, periodically crunching the FFTs and delivering it to each.

Will the GPU always be faster than the xN CPUs?

There are some major memory bandwidth problems getting the data up into to the GPU memory and back down again.

Certainly for the asymmetric bandwidth of AGP. Still a problem for PCI-E?

Once I break FFT out into its own class with versions for each current FFT method (Intel, FFTW, Ooura) it should be easier to add GPU FFT as a test situation.

Now that gets a WOW!

Is this going to be for the present piplines GPU systems or only for the new CUDA stuff? Or both? (Or just the GPU on your test system? :-) )


This sounds seriously interesting.


Happy crunchin',

Regards,
Martin

---

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)

Surely the problems of using the GPU are;
1, most modern pc's using pci-e have only one gpu but multiple cpu's
2, and although pci-e is fast, compared with previous interfaces, it communicates at max 2.5 Gbit/s per line, but there are 16 lines to these gpu's. But compared to main memory bandwith of my Pent M that is showing 8.3 GByte/sec, but presume my e6600 is even faster.

Andy

Nokia is going to make SDK available "to those whom it should go to". I hope that includes Seti@home!

Nokia? That sounds like a cell phone maker.

Don't You mean Nvidia?

Of course, a typo. Nvidia.

Srry

---

Not to bring a dead thread back to life, but if anyone is attempting to create code under CUDA, I have an 8800 GTX and would be willing to test as needed.

---

Not to bring a dead thread back to life, but if anyone is attempting to create code under CUDA, I have an 8800 GTX and would be willing to test as needed.

same here...

---

The development kits for CUDA were released to the public...hopefully this will help in getting a Seti app optimized for us 8800 owners so we can rock the RAC.

---

This thread also interests me. I have one of the new Mac Pro machines so I have 4 * 2GHz Intel 5130 Xeon cores as well as an Nvidia 7300GT.

Cheers,

MarkL

---

8800 GTS driving the monitor on my E6700 based machine. Given that the majority of the time it is in screen saver mode (with the monitor blanked), there's a LOT of spare machine cycles just waiting to be utilized. :-)

---

for those that are interested there is are GPU tests for FFT and FFT 2D available on Simon's site.

AFAIK there have not been any 8800's tested.

thread link http://lunatics.at/index.php/topic,152.0.html

-citroja

---

for those that are interested there is are GPU tests for FFT and FFT 2D available on Simon's site.

AFAIK there have not been any 8800's tested.

thread link http://lunatics.at/index.php/topic,152.0.html

-citroja

Well If I can find any links there will be an 8800 tested....

I have been waiting activation code to lunatics forum few days now,but looks like it wont come.... I would have 8800GTX in my puter.

---

WR,

some people seem not to get the activation e-mail because of over-zealous spam filters.

So please, email simon |at| zadra <dot> org with your user name and I'll help you get it sorted.

Thanks,
Simon.

---

Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

Surely the problems of using the GPU are;
1, most modern pc's using pci-e have only one gpu but multiple cpu's
2, and although pci-e is fast, compared with previous interfaces, it communicates at max 2.5 Gbit/s per line, but there are 16 lines to these gpu's. But compared to main memory bandwith of my Pent M that is showing 8.3 GByte/sec, but presume my e6600 is even faster.

Andy

At folding@home 1 gpu = about 30-40 CPU's (estimated)
So there is definatly some benifit.

BoB

F@H stats

---

Do you Good Search for Seti@Home? http://www.goodsearch.com/?charityid=888957
Or Good Shop? http://www.goodshop.com/?charityid=888957

Hi,
According to ars technica http://arstechnica.com/news.ars/post/20070227-8931.html and Bryan O'Sullivan and what I understand of it, we have to wait a long time before we can put our Gpu's to work.

My guess, 1 - 2 year for a skilled and dedicated programmer, and it could well be more - a lot more. Normaly you need a company to do this kind of programing. If it's not your job but your free time, nha.. it's too consuming.

---