ATI folding

[Nathan_P]

Hi everyone.

i know it gets asked occasionally. what is the current issue with ATI and folding? Well over at FF there is a long thread about support for the 5xxx cards and why is it taking Pande so long to get organised. Since Vijay has offered an explanation and a challenge i thought everyone would like the chance to read.

5xxx support thread:>http://foldingforum.org/viewtopic.php?f=51&t=13647

Vijay's challenge post:>http://foldingforum.org/viewtopic.php?f=51&t=13647&start=60#p144900

 

[Nessus85]

Wow. What I took out of that was that they were just "making it work" as opposed to optimizing it. And at that, it seems they are only doing it when they have spare time. I thought for sure they would want the extra ooomph that would be provided from ATI users Hopefully someone steps up to the plate...................but I wont keep my hopes up.

 

[zero2dash]

(Without reading the articles)
From what I've read in the past, it sounds to me like it's a bunch of finger pointing. Pande blames ATi, ATi blames Pande. Pande seems more to blame for using an un-optimized client. ? That seems like ATi's stance. Pande seems like they're saying ATi isn't helping. ?

I don't know who to believe.

 

[Nathan_P]

Wow. What I took out of that was that they were just "making it work" as opposed to optimizing it. And at that, it seems they are only doing it when they have spare time. I thought for sure they would want the extra ooomph that would be provided from ATI users Hopefully someone steps up to the plate...................but I wont keep my hopes up.

Its more a question of resources than anything else. They have had to do a lot of work over the whole of the project recently in order to keep up with the amount of science that they want doing and the amount of donors they have to feed with WU

New protomol core for CPU's, new SMP2 client and core, new server code as cross all the servers, new gpu clients and cores for both ATI & Nvidia plus actually doing some science. Given this i suppose they went into "lets get the easy stuff done first" mode and will now start on the ati cores.
The other thing to consider is that at least GPU2 works on all ATI cards, nvidia broke GPU2 when they went to fermi and all those cards were doing no work at all.
I've learned over the last couple of years that with stanford patience is a big virtue, they will get the work done, but only they know how far along they are and a likely release date

 

[sirmonkey1985]

(Without reading the articles)
From what I've read in the past, it sounds to me like it's a bunch of finger pointing. Pande blames ATi, ATi blames Pande. Pande seems more to blame for using an un-optimized client. ? That seems like ATi's stance. Pande seems like they're saying ATi isn't helping. ?

I don't know who to believe.

and yet the funny thing in all of this is that ATI cards are used for the benchmarking and testing of WU's.. so why use a card brand to create your bechmark numbers if you arent going to give a shit about optimizing the client for that card..

 

[Jon855]

Wow. What I took out of that was that they were just "making it work" as opposed to optimizing it. And at that, it seems they are only doing it when they have spare time. I thought for sure they would want the extra ooomph that would be provided from ATI users Hopefully someone steps up to the plate...................but I wont keep my hopes up.

Me either - I'm not gonna keep my hopes up - they're just dicking around that's for sure.

(Without reading the articles)
From what I've read in the past, it sounds to me like it's a bunch of finger pointing. Pande blames ATi, ATi blames Pande. Pande seems more to blame for using an un-optimized client. ? That seems like ATi's stance. Pande seems like they're saying ATi isn't helping. ?

I don't know who to believe.

Don't believe them - do you really think AMD will be like our GPUs can't do these sort of thing and be all broken? Nope.

and yet the funny thing in all of this is that ATI cards are used for the benchmarking and testing of WU's.. so why use a card brand to create your bechmark numbers if you arent going to give a shit about optimizing the client for that card..

Exactly - this is stupid - really.

 

[10e]

Correct me if I'm wrong, but I was under the strong impression from the past that the ATI cards were great at doing far larger protein molecules than NV cards, and that this would be a valuable contribution to the project.

While maybe ATI's developer platform and relations lag behind NVidia's, I find it hard to believe that it's been nearly two years since there were any new developments with ATI folding.

To me it reeks of politics, plain and simple.

We'll see what happens when OpenCL becomes a standard and OpenMM supports it, but I'm not holding my breath.

 

[ghost6303]

according to those links, ATI cards are limited to using 300 shaders at a time currently. so for those of you with 5870's with 1600 shaders, only a fraction of the card is being utilized, which is where the performance hit comes from.

they only have one full time and one part time employee dedicated to all the GPU coding work. its not like there is a team of dozens of people working on this- just one and a half people. and that is encompassing all of GPU folding work, not just ATI. you cant blame them if getting the nvidia 4xx series of cards working was an easier task so they picked that first.

ATI uses their own low-level programming language for their cards (CAL -compute abstraction layer) that the OpenMM software, which does the folding computations, does not easily or directly port into and is not compatible with nvidia, so it must be almost entirely rewritten instead of just repackaged. they also use their own optimization programming which is non-standard.

they arent just sitting on their thumbs with nothing to do here. i want an ATI client as much as the rest of you, but you cant sit here and bash them for not handing it to you, for free, sooner. as said on the forum- if any of you want to write a CAL platform of OpenMM, the source code and developer tools are all open and freely available on the website. go nuts.

 

[Mr. Pedantic]

according to those links, ATI cards are limited to using 300 shaders at a time currently. so for those of you with 5870's with 1600 shaders, only a fraction of the card is being utilized, which is where the performance hit comes from.

This kind of makes sense, seeing as performance hasn't actually improved from the 4890 to the 5870 (which have the same clocks, from what I remember).

they only have one full time and one part time employee dedicated to all the GPU coding work. its not like there is a team of dozens of people working on this- just one and a half people. and that is encompassing all of GPU folding work, not just ATI. you cant blame them if getting the nvidia 4xx series of cards working was an easier task so they picked that first.

They shouldn't be doing things first just because it's easy. That's definitely not how I study, and it's definitely not how I'd do my research.

i want an ATI client as much as the rest of you, but you cant sit here and bash them for not handing it to you, for free, sooner

To be honest, if they don't it's their loss. Nobody else's.

To me it reeks of politics, plain and simple.

cf. TWIMTBP.

 

[ghost6303]

They shouldn't be doing things first just because it's easy. That's definitely not how I study, and it's definitely not how I'd do my research.

well the guys doing the coding arent the ones who are interested in the results of folding. and if hypothetically you have a 50/50 split of ATI and nvidia users, and its 5 times as easy to get the nvidia cards working as ATI, I would certainly make that first priority.

To be honest, if they don't it's their loss. Nobody else's.

very true. except maybe some points for [H]. EVGA's ppd is made up of almost entirely nvidia cards, while ours has many ATI cards mixed in. fixing the ATI client would give us more of a boost. but aside from loosing out on fictitious internet points, its a loss to science.

 

[amdgamer]

well the guys doing the coding arent the ones who are interested in the results of folding. and if hypothetically you have a 50/50 split of ATI and nvidia users, and its 5 times as easy to get the nvidia cards working as ATI, I would certainly make that first priority.


very true. except maybe some points for [H]. EVGA's ppd is made up of almost entirely nvidia cards, while ours has many ATI cards mixed in. fixing the ATI client would give us more of a boost. but aside from loosing out on fictitious internet points, its a loss to science.

This last part is a good reason to motivate us to hold on as long as we can. As others have said, this does definately reek of politics and special interests. How long can they continue to ignore the ATI crowd?

 

[ghost6303]

Im still not sure where the evidence that they are specifically snubbing the ATI crowd is. lots of people keep saying this, but im not exactly sure why or what they are basing this off of.

to quote this again

To be honest, if they don't it's their loss. Nobody else's.

they are only shooting themselves in the foot if they artificially hold back on support for any set of hardware. i see zero motivation for them to do this.

also directly from the horses mouth over at stanford, and this is true- look it up if you dont think so; they have repeatedly said CUDA is easier to implement the molecular modeling code on.

CUDA has been around for longer (since geforce-8 series in 2006), and as such, is better supported with a larger code example library and people who know how to write for it, and is more implementation friendly then CAL.

CAL is a much newer language from ATI (first released with their SDK, around 1/2008) which trades extra complexity and stricter code structure for more utility and more powerful features. it does not port nearly as easily with the OpenMM code. and there are not nearly as many coders out there with experience with CAL as there are those for CUDA.

i am unaware of any noteworthy monetary or hardware or software or other donations from either hardware vendor to stanford that would back a "political bias" claim. i have never seen any press release, forum post, news, anything, that would lead me to believe that this delay was caused on purpose, however unfortunate it is to us all. if anyone wants to show some, feel free. otherwise claiming stanford has some sort of "political bias" towards some hardware vendor based on nothing does nobody any good.

 

[Nathan_P]

^^^This

 

[ghost6303]

good links btw, Nathan.

 

[Jathanis]

<snip> but aside from loosing out on fictitious internet points, its a loss to science.

Not necessarily - my two nVidia cards still fold at 7K ppd give or take, but my ATI 4850 has gotten me about 13M points over on science projects like milkyway@home and Collatz. I figured I might as well get the best bang for my bucks out of the cards. If a ATI GPU client that fully utilizes my card comes along, my ATI card will probably join my other 2 nVidia (soon to be 3, woot) folding, but until then, I'll keep tham maxed out on other science projects! Only loss here is to F@H, but with all the SMP defectors some time ago, myself included, we can conclude that this isn't the first or the last time this will happen with Stanford...

 

[Mr. Pedantic]

very true. except maybe some points for [H]. EVGA's ppd is made up of almost entirely nvidia cards, while ours has many ATI cards mixed in. fixing the ATI client would give us more of a boost. but aside from loosing out on fictitious internet points, its a loss to science.

not quite what I meant, but okay....

 

[Nathan_P]

good links btw, Nathan.

No worries, there has been a long and somethimes heated debate over at FF regarding this and seeing as it pops up here on occasion and a lot of people here use ati gear I thought it good to pass along.

 

[ccityinstaller]

Not to add gas to the fire, but the stupidly low PPD on my 4850 is the reason I just swapped it for an EVGA 8800 GTS! The power draw vs PPD on the 4850 was simply unsustainable. I pay between 150-200 a month for electric, and this is with only 1 desktop running 24/7.