Latest discussions at Einstein made it necessary te recheck my sight of things.
The performance chart I was referring showes the performance with cuda32 apps.

Alex, the reason for Einstein using CUDA32 (according to Bernd) is:

"Our BRP application triggers a bug in CUDA versions 4 & 5 that has been reported to NVida. The bug was confirmed but not fixed yet. Until it has, we're stuck with CUDA 3."

http://einstein.phys.uwm.edu/forum_thread.php?id=10028&nowrap=true#123838

He goes on to say they reported the bug to NVidia 2.5 years ago. Still not fixed. Glad we didn't hold our breath...

Alex, the reason for Einstein using CUDA32 (according to Bernd) is:

"Our BRP application triggers a bug in CUDA versions 4 & 5 that has been reported to NVida. The bug was confirmed but not fixed yet. Until it has, we're stuck with CUDA 3."

http://einstein.phys.uwm.edu/forum_thread.php?id=10028&nowrap=true#123838

He goes on to say they reported the bug to NVidia 2.5 years ago. Still not fixed. Glad we didn't hold our breath...

Yes, thanks for the link, I know this thread. But I did not know that there is a performance increase of about 2 (according to the discussion at Einstein) when using newest hardware and cuda 4x.

I have no way to compare the performance of cuda 4x wu's against AMD OpenCL. But with this (for me) new information I have a feeling of some kind of unfairness compairing cuda3x against newest AMD OpenCL. This is what I wanted to point out.

As far as the developement at Einstein is concerned: They do a great job there, and they have good reasons not to use cuda4x. And they have wu's for Intel GPU and also for Android (at Albert).

I have no way to compare the performance of cuda 4x wu's against AMD OpenCL. But with this (for me) new information I have a feeling of some kind of unfairness compairing cuda3x against newest AMD OpenCL. This is what I wanted to point out.

All you can compare is what's there. If NV won't fix their bugs, that's their problem. You can certainly compare OpenCL on both platforms in various projects. Suposedly that's where GPU computing is going...

GPUGrid has seen plenty of CUDA problems in the distant and recent past, and there are reasons why GPUGrid didn't move from CUDA4.2 to CUDA5 and have problems with the existing 4.2. There are different problems at different projects - some can adapt, some can't.

Porting CUDA to OpenCL would result in a performance drop for NVidia cards and with anything other than GCN performance would be far too poor. Even GCN is buggy and anything but the most simple models is still very inefficient.

BTW OpenCL isn't limited to GPU's, it can run on CPU's. Forthcoming APU architectures from AMD should prove useful for some research techniques. Certainly an opportunity exists to explore this potential, but whether it will suit this projects research ambitions or not is down to Gianni.
In my opinion new cures and treatment strategies can only be facilitated by new research techniques and avenues, but ask in 6 or 7 months when the technology arrives. Note that it's not all about how fast something can be done, it's also about what's new that can now be done that previously couldn't; GPU's aren't just getting faster, they also introduce new research resources.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Im (im)patiently waiting for GPUgrid to get their app working so I can bring my 780 here. Currently it is at Einstein. While I love the project, for what it is, the fact that they're effectively stuck on an older CUDA platform, thus not getting the most they can out of the newer cards makes me not want to crunch there.

It's most definitely a GCN world over there. Yes, you can crunch just fine there with NVIDIA GPUs, just don't expect any performance increases through different app updates.

If it weren't for GPUgrid, I would probably have an ATI rig. I still want to build one, problem is, I like building high end machines. I would rather all my resources go hear. :P

Different cards for different purposes I suppose is what it really all boils down to though. And there's nothing wrong with that.

Even GCN is buggy and anything but the most simple models is still very inefficient.

Anything complex enough to be called a processor is buggy. We need firmware and drivers to work around these issue. And to make sure they don't get too far in the way of actually using the chip.

Regarding GCN not being efficient for handling complex code: I'd say it's actually the other way around, that nVidia took a step back with Kepler for complex and variable code (while gaining efficiency for regular and simpler cases). If things are normal, both architectures are about on par. But there are certainly cases where GCN completely destroys Kepler (e.g. Anand's compute benchmarks). This could be due to bad OpenCL drivers rather than hardware.. but then Fermi shouldn't fare as well as it does in comparison, should it?

And there are of course cases where nVidia running CUDA destroys AMD running OpenCL.. software optimization is a huge part of GPU performance.

Anyway, the GPU-Grid team seems to have their hands full currently with a reduced staff and many problems popping up. Not much room for further development when bug fixing is so urgent...

MrS

---

Scanning for our furry friends since Jan 2002

We ended up designing and building our own GPU chassis so tired of having poor cooling.

Ooh, show us pictures! I want to build a solid (silent) cool (as in temperature, not epicness) case for multiple computers (to make a central house computer, much like central heating, since it will also heat the house), and need ideas.

Here you go http://www.acellera.com/products/metrocubo/

There is no way I would buy a Blue case
( ; P

They actually make them in any color :D We have a green and an orange one in the lab and I am still waiting for my pink one, hahah

and I am still waiting for my pink one, hahah

Isn't that Noelias workstation? ;)

MrS

---

Scanning for our furry friends since Jan 2002