My comments on the Folding@Home

[filu]

Recently I'm very frustrated, what is happening with the project Folding@Home. There are several reasons why I am considering leaving the project. Here are a few of them.
1. Permanent beta clients. Immediately raises the question whether the results of the folding are for something useful? In my opinion client's beta phase, with such major projects should not go beyond the university and a group of beta testers.
2. Too large disparities in scoring each of the WU example GTX480 is at least twice more effective than a GTX285, and the PPD is only 50% higher.
3. Why are WU's on Uniprocessor client have a cut PPD? But the main principle of the project was to use to fold unused CPU power. This "clever" move resulting in an outflow of ordinary people.
4. I have three dedicated computers for folding bigWU, and they are over 50% of the time to fold projects 6701, 6702 or 6024. Frankly the difference in PPD between projects 2685, 2686 or 2692 and these are too large: 20 KPPD. Lack is here any logic when determining scoring individual WU, like when scoring WU's for ATI and NV.
This is not an attack on the project, but my observations that PG might think. Probably, if it bore the Stanford University total costs associated with the project idea of the project would be more thought out.

[Ivoshiee]

Permanent beta status is more a philisophical matter. It either does work as intended or not. At the same time I am with you what goes into constant tweaking and "caring" of the FAH client. That is not a good thing. Install and forget should it be.
What goes to WU scoring then the scoring system is supposed to be KISS. Is it or isn't it or does it produce "a fair" results for the donors, it is hard to tell.
Scoring system intent is to direct donor system-client decision to FAH favourable direction. It should, but does it do it?

[k1wi]

Hi filu,

From what I have learnt from the programme, I could be wrong in some instances, but I'm sure a mod will correct me where I am:

1. The permanent beta clients - The SMP client is a high performance client because it simply isn't as rock solid as the uniprocessor client. It isn't permanently in beta, though when new versions are released they are in a public beta, just like Windows7 was for the better part of 9 months! This is primarily because it is absolutely pushing the very boundaries of SMP computation. The cost of continuously pushing these limits is that it isn't rock solid 100% of the time. For those users who don't want to run that risk, there is set and forget uniproc client. Users accept the risk and are warned of the risk. If it was just left within the university there would be a significant slow down in progress. (Just for the record, work units are thoroughly tested prior to release, but sometimes things don't make themselves known until they're 'in the wild' across a whole range of configurations). GPU3 is beta because it is reallly reallly a moving target as it's developing a field that is still in its infancy.

2. On what basis do you consider the 480 to be twice as effective as a 285? peak theoretical FLOPS? Do remember that the internal design of a GPU (or CPU as I'll mention later) has a major impact on Folding performance, it could be that something in the design of the 480 means it doesn't perform as for FAH. What do you do? Rewrite the core so that it works on the 480, at the expense of the 285?

3. I can't answer that, other than to say that my own opinion that the higher risk, higher bandwidth, higher memory demands of the SMP client mean they should get a better return, just as the BigAdv does.

4. Thank you for your contribution, 3 dedicated machines is a big investment in the project. The work units are all benchmarked on a standard machine. But I suspect you don't have the standard machine. It could be that your computer has a significantly larger amount of cache, or perhaps less. It is feasible that, that extra amount of cache means that one project folds much more effectively on your machine than another one does. That's not a fault of benchmarking, it's just repeating point 2 that different hardware configurations can have different effects on performance, across projects and you can't benchmark every project for every computer!

4a. The ATI NVidia Chestnut! There are a number of really really good posts outlining how the architecture between the two companies differ, and how this has a major effect on relative performance, along with how the development of the different languages have hindered/encouraged development.

Finally, I would hasten to add that there is the publicised V7 coming up that sounds like it'll be a big step forward - which is what FAH is all about - pushing the boundaries and making progress. I hope you stick around.

[filu]

2. On what basis do you consider the 480 to be twice as effective as a 285? peak theoretical FLOPS? Do remember that the internal design of a GPU (or CPU as I'll mention later) has a major impact on Folding performance, it could be that something in the design of the 480 means it doesn't perform as for FAH. What do you do? Rewrite the core so that it works on the 480, at the expense of the 285?

http://www.anandtech.com/show/2977/nvid ... he-wait-/6

Finally, I would hasten to add that there is the publicised V7 coming up that sounds like it'll be a big step forward - which is what FAH is all about - pushing the boundaries and making progress. I hope you stick around.

In my opinion
Client V7 = Duke Nukem Forever

4a. The ATI NVidia Chestnut! There are a number of really really good posts outlining how the architecture between the two companies differ, and how this has a major effect on relative performance, along with how the development of the different languages have hindered/encouraged development.

I write only in this way. In other DC projects tackled this problem. I will not mention the projects, because I'll stay again accused of recruiting. But maybe this is causing difficulties?
http://en.wikipedia.org/wiki/Jen-Hsun_Huang

[7im]

I write only in this way. In other DC projects tackled this problem. I will not mention the projects, because I'll stay again accused of recruiting. But maybe this is causing difficulties?
http://en.wikipedia.org/wiki/Jen-Hsun_Huang

Suggesting a conspiracy is way off base. You have no supporting evidence, nor did you offer any. If Jen had anything to do with GPU and Folding, then WHY was the very first GPU client an ATI ONLY CLIENT?!!! And why was it was the only GPU client for 2 years?

Give me a break. NV has the better hardware and software for folding right now. Plain and simple.


And clients are not permanent beta. The Linux SMP client is no longer beta. And the v6.30 SMP Windows client beta was a big step forward. The new bonus points system is another big step forward.

The point system as been debated many times at length, please read the Points FAQ as a start. It will answer your points questions. Clearly your observations are biased unless you understand the whole system better. Points are based on scientific production. Single core CPU clients produce very little science compared to smoking fast GPUs. And more work done equals more points. Why isn't that fair?

The uniprocessor client is still benchmarked at 110 PPD on a 2.8 GHz Pentium 4. Please provide evidence that that value has been reduced.

[filu]

he uniprocessor client is still benchmarked at 110 PPD on a 2.8 GHz Pentium 4. Please provide evidence that that value has been reduced.

The Linux SMP client is no longer beta. And the v6.30 SMP Windows client beta was a big step forward. The new bonus points system is another big step forward.

And what about bigWU for Linux? Sure everything is okay, just me picking up this?

[PantherX]

Regarding bigadv, it is still a trial thus it isn't a "standard" category yet.

[bruce]

The Linux client is no longer beta. The Linux version of FahCore_a3 does not work correctly with bigWU.

Perhaps the distinction isn't important to you, but the actual Stanford client and Stanford servers can downloading WUs, invoke a viable FahCore, upload the results and credit points.. Gromacs.org has been correcting bugs in their code and Stanford will create a new version of FahCore_a3 from that code "soon", but they're not going to do that until the core produces reliable scientific results when the WU exceeds a certain size.

[Tobit]

The uniprocessor client is still benchmarked at 110 PPD on a 2.8 GHz Pentium 4. Please provide evidence that that value has been reduced.

Sounds about right as my 3.0GHz P4 does 180 PPD. OP should have been folding when the standard client was all that was available, maybe he'd appreciate things differently. Although some of the GROMACS code has been made more efficient over the years, uniprocessor technology hasn't changed, why should the points system? Uniprocessor work has long deadlines so they can be returned more leisurely. Stanford doesn't need to get the results back as quickly as they do with the SMP client thus the lack of the need for bonus points.

[7up1n3]

Give me a break. NV has the better hardware and software for folding right now. Plain and simple.

Software, yes. Hardware? Not so much.

And clients are not permanent beta.

WinSMP has been in beta since its inception. I think the point being made is that we've had clients (or series of clients) that remain in the beta stage for years - which may feel "permanent" if you're not a graybeard like me.

Yes, v6.30 is a huge step ahead and IMHO is the first WinSMP client that could have the beta label removed. And, again IMHO, it should be removed - because its discouraging users from giving it a shot, because its still labeled "beta" despite being light-years ahead of previous WinSMP beta clients.

[7im]

v6.30 did have the beta label removed (in spirit), except for the fact that client is still beta testing a new SMP2 fahcore (a3), so it kept the beta label.

It's more of a "preview" client than a "beta" client.

SMP2 preview client (v 6.30)

[k1wi]

I strongly recommend anyone who wants to complain about ATI performance vs. NVidia performance and say hardware doesn't cause anything to read the post below:
http://foldingforum.org/viewtopic.php?p=117995#p117995

You can't call conspiracy if you haven't first read that post. The fact remains Hardware design plays a huge factor. Combine that with changing firmware and software and you pretty much have the current state. I've seen it written around, but FAH makes very complicated calculations, more complicated than some of the other 'projects that seem to take full utilisation of the GPU', so it is much more sensitive to hardware design changes and firmware/software changes.

[7im]

Oh, and if there is any doubt the v6.30 client is a great improvement, read this thread...

Love the v6.30 SMP client!

[mdk777]

I strongly recommend anyone who wants to complain about ATI performance vs. NVidia performance and say hardware doesn't cause anything to read the post below:
viewtopic.php?p=117995#p117995

You can't call conspiracy if you haven't first read that post.

No conspiracy, just a lack of effort in support.
The conclusion to the thread you reference is:

As has been explained ad infinitum in this thread and others, our focus is now on the next generation GPU3 client and OpenMM core. The programming language and model used by the old ATI code have been made largely obsolete by the introduction of OpenCL, so we're focusing our efforts there, rather than trying to revamp an older code.

Here is clearly admits that the client was for all intents and purpose depreciated a long time ago. It just was never made official.

Yes, this is all several year old news, but continuing to claim it didn't happen falls into the category of history revision.

[k1wi]

It's not depreciated - there is still work units available for it. But rather than continue to develop/improve something that is a sunset path, they're focusing on developing the future. Pande have made no bones about developing the future being a LONG path, but one that they are calculating, in the long run, will pay off better