beta 8

[Brucifer]

So now we have beta 8 and this is for cuda 2.2. And like the beta 7 cuda 2.2, this beta 8 is slow. No beta 8 cuda 2.1. Running the -bench command and I'm showing about a 33% speed loss on both the 9800GT and 260GTX.

And I also didn't have over 10,000 status units show up in the accounts they were crunched in for yesterday's stats run (27th). So I don't know what to say.... other than I'm really rather dismayed at how this project is turning out. Seems to be losing ground rather than gaining anything.

and on the nvida site, I see that cuda 2.3 windows driver is out, so that means more beta's will be coming down the road.

Seems like an old addage is appropriate;

"When you are up to your ass in alligator's it's hard to remember that the initial objective was to drain the swamp."


So maybe it's just time to move away from the swamp. :icon_rolleyes:

[NeoGen]

Just thinking out loud... I think I can understand why they want to move to higher versions of the cuda library, most likely there's nasty bugs in the older ones, and new features in the new ones that interest them. It's usually what happens when working with a library that is still being actively developed.

And then there's also the speed vs precision tradeoff, if you need more precision in the calculations you'll have to sacrifice speed to get it. I have a feeling that's what happened, since there was such a noticeable performance decrease.

But in the end, I really have no clue on what happened with those guys at dnet. I'm just thinking out loud here. :)

[Brucifer]

After a night of running the new beta 8 client I'm pretty nonplussed about it. Terrible output compared to the previous betas.

Not picking a fight so don't take it that way neo. If precision was/is the issue, then why are they retaining the previous completed gpu work? That would lead the man in the street to assume that there isn't any problem with the quality of the completed work. I can understand wanting to move to updated toolkits as they generally feature more complete/robust routines. But from the performance issue, this one slides backwards. To gain more output, then one has to buy newer, faster hardware. Sorta like windows code, the more goo-gaws added to windows, the more powerful the processor required to run it and keep response snappy. Same thing here with rc5. So why put all this into a project with a known answer, with the only benefit for some person/team of the thousand dollar prize. About the only benefit of note I see here is that some guys/gals writing gpu code for rc5 are getting experience.

my argument basically sits that if earlier gpu betas were good enough for the results to be accepted into the permanent results database, then that code that produced more results was the more beneficial towards completion of the project. Granted the code may not be as "elegant" as the latest code, but the drive towards using the gpu was to try and significantly speed up the possible completion of the project. In this regard things are slowing down rather than speeding up.

So I look at this and where it's been and looks like it's heading, and then I look at what folding@home is accomplishing and it's becoming more and more clear to my thick skull that it's time to move on to folding@home or one of the others. yeah yeah yeah, I know, :-) I've said that before a few times. But this has become frustrating and is not an enjoyable ride anymore.

[liuqyn]

let me play devil's advocate for a minute, maybe the folks at distributed.net are just using RC5 as a GPU test bed for a future application for OGR which would be more beneficial and arguably require greater precision as well.

just a thought.

[Brucifer]

They've already said that the OGR algorithm doesn't lend itself to the required format for crunching by cuda. They've already looked at it and it's a dead-issue.

As for experience for a future project, that's an entirely possible thing for sure.

[NeoGen]

Yea, that would be nice. They could try the old RSA challenges for example. Altough I remember that a few years ago RSA closed down the contest, the challenge numbers are still out there and unsolved. :)

[Brucifer]

I see in the bugs thread that the subject of the 1/2 speed on the cuda 2.2 client has been brought up.

[Brucifer]

It seems as though I'm not getting any points submitted from Beta 7 which is not expired yet, which I stepped back to as Beta 8 runs at half speed. So that makes two days crunching that disappeared into the big void in the sky. So I've decided to pull the plug on rc5 altogether. Just tired of the games, better things to spend my time and money on.

[Brucifer]

ah.... the confusion clears......

http://n0cgi.distributed.net/cgi/dnet-finger.cgi?user=bovine

So it looks like a bunch of work by the betas prior to the cuda 2.2 is flawed... :-( That's too bad, both for those that crunched, but more so for the project as there is some work that will have to be recrunched at some point if they get that figured out. And it also answers why some of my work fell off the face of the planet.

So we now be crunching at half the previous speed with the cuda stuff. While that sucks, it's still a much faster than the x86 clients with the plus that we now know that the work completed is completed correctly. I'm assuming now that they really gave the results a good going over to make sure there weren't any more bogus units being returned by the cude 2.2 beta clients.

[liuqyn]

very informative. still not sure I'll play with it again though.

[Brucifer]

It's good they caught the problem before things got too far down the road. But that's why it's beta.... :)

I will probably still play with it some, but not like I used to. Just one of those things that time will tell. I always like the distributed.net clients in the past as they were stable and all. And rc5 has been too with the exception of the learning curve with cuda. And as it turns out since it was a compiler issue, it falls under the growing pains thing. It has and will continue to be a good experience for those programmers working with cuda. I'm just happy to finally know why stuff was falling into the big void.

As for the answer being known... it is always nice to know something for sure. It's the challenge of it I guess that keeps people crunching away on it. There have been a lot of projects go down the tubes in the recent past. So there isn't any guarantees that any of this stuff that any of us are working on will really pan out for much of anything. Some projects have been around for a long time and have a reputation established on completed results. Others have inferred that they have results, but nothing that the man in the street could really lay a finger on. So in that regard, distributed.net has been a pretty solid project management effort over the long haul.

[Brucifer]

Got back from a trip just in time to see the cuda beta8 expire, and once again no replacement ready to go. Not like this is the first time it's happened or that I've commented about it. However things just aren't the same anymore. I don't enjoy this stuff after spending bucks on hardware and electricity. So what you see at the stats run tonight team is what you get and what you've got. Time to invest time and effort into something else in life.