News from PG re GPU benchmarking

they are probably doing that

they could also have a standard SMP WU they ported to GPU (most likely)
could have ran it internally(on both) and balance the scheme according to their benchmark machines

once they get comparisons back on tpf and points they can go back to the SMP points & tpf for comparison

you have no Idea what they have done and you are conjecturing
like everyone has said let this play out ...
stanford will look at the results and determine what's right... not you, your team, me or my team

let beta testers give feedback and enjoy the points that come from the odd GPU beta WU you get

He will have more idea than us - Kendrak is a DAB rep, not just a team member, Oh and last time i checked over at FF beta testers could not give feedback because everyone else was stealing the WU, and since they are not on the beta team they cannot discuss their findings.
 
The biggest thing is the implicit vs explicit factor. I have a hard time trying to understand how a gpu can work like a 4P rig.

Maybe they have a new use for smp/bigadv work...
 
I'm currently chewing on a 6947 WU with 77,113 atoms on a pair of L5640's with a tpf of 1:52 for 42k PPD. Last time i checked a GPU WU topped out at about 2000 atoms and everyone was bitchin about the heat, power draw and tpf when the got one.

Feed a 570 that 6947 WU and see how it does, if its faster then fine it earns more points, if its slower it earns less. I don't have a problem as the work will be the same. Personally i think the problem is moot as the GPU will melt trying to to process the info
We've tried that last weekend. FahCore_15 barfs on all SMP units; looks like we need to
compile GROMACS/OPENMM to measure. Though I did the opposite -- ran GPU unit on
couple SMP boxes.

The issue at hand is more convoluted; I'll be posting (sometime in the evening) detailed
analysis and possible consequences/ramifications to the community.
 
He will have more idea than us - Kendrak is a DAB rep, not just a team member, Oh and last time i checked over at FF beta testers could not give feedback because everyone else was stealing the WU, and since they are not on the beta team they cannot discuss their findings.


I mentioned this earlier
let the Beta testers do them.
not a good idea for everyone but the cats meow to jump on them when they are so rare at the moment

I sure haven't gotten one and Im not configuring all my rigs to get them either

and yes they beta members can discuss the findings on the ff beta forum...just not here
you do have the ability to go read the post there and see what they post

kendrak is doing a great job for you guys. Im sure every rep will have something to say

the GPU QRB bonus has been work on for long time and was discussed at the DAB at length.
Stanford would not have release the units and GPU QRB if they weren't confident in their findings
 
I am yet to see evidence of QRB for GPU not having been done ad-hoc.
 
I am yet to see evidence of QRB for GPU not having been done ad-hoc.

^^This, in fact IIRC this 900 atom WU is reminisent of the fermi test WU that was released when Fermi was 1st released. Lots of people bought 450GTS cards only to realise that they choked on a proper fermi WU.

If PG are serious about " equal points for equal work" they must feed a GPU a SMP sized WU and see what it does asnd how the GPU copes
 
We've tried that last weekend. FahCore_15 barfs on all SMP units; looks like we need to
compile GROMACS/OPENMM to measure. Though I did the opposite -- ran GPU unit on
couple SMP boxes.

The issue at hand is more convoluted; I'll be posting (sometime in the evening) detailed
analysis and possible consequences/ramifications to the community.

tear like a bawss. :cool:
 
He will have more idea than us - Kendrak is a DAB rep, not just a team member

The reason I say it is all a guessing game (at least as far as the DAB is concerned) is that VJ posted a draf of the blog post and says (to effect) QRB is here for GPU. Look this over, I will post it soon.

We gave thumbs up and it was posted.

Beyond that nothing technical has been posted or talked about in the DAB.
 
I'm seeing about 96K ppd on a gtx 570 w/ the "new" 8057 WU. Relying on the FAHControl app for the ppd estimate right now so I dunno how accurate that is. TPF is 2 min 29 secs.
 
460s, stock clocks, now running WU 8057.
TPF of 3:31. PPD of 57,131
The TPF on the 560Ti for WU 8057 is 2:15.
 
That's interesting a 560Ti being faster than a 570? Is that 560Ti one of the 448 cuda core variety?
 
I am yet to see evidence of QRB for GPU not having been done ad-hoc.

This is the issue.

Truly equal pay for equal work isn't a problem, but this odd Wu selection doesn't feel right. If pg wants to disclose their thought processes more, we are open to dialogue.

Having all this "helpful" input from other teams isn't needed or constructive at this point.
 
That points to higher clock speeds being more important than cuda core count which I guess isn't all that surprising.
 
That points to higher clock speeds being more important than cuda core count which I guess isn't all that surprising.

That would be a change, PG has always preferred more cores in the past
 
Truly equal pay for equal work isn't a problem, but this odd Wu selection doesn't feel right. If pg wants to disclose their thought processes more, we are open to dialogue.

However, the Great Panda is not open to dialogue. They have never been open to dialogue or reason.
 
However, the Great Panda is not open to dialogue. They have never been open to dialogue or reason.
There should be two sides to running f@h, the important one that takes care of the science, and the fun and games team who keeps the PPD competitive. Pande says "we have a new WU to run" and gives it to the points team. Points team benches it and sets points to benefit new hardware, maintain competition, and incentivize investment in a particular type of hardware.
 
There should be two sides to running f@h, the important one that takes care of the science, and the fun and games team who keeps the PPD competitive. Pande says "we have a new WU to run" and gives it to the points team. Points team benches it and sets points to benefit new hardware, maintain competition, and incentivize investment in a particular type of hardware.

Wait, What!? Folding for fun and science!?

Madness!!!1!1 :D
 
Guys lets get back on topic for the New units, not on the dialog, or lack there of, from PG
 
PG might have finally found a way to get me to upgrade _some_ of my clients to 7 then.
 
There's more to a WU than just atom count, there's also steps involved. With the initial bonus of 10x,
I'd say that GPU folders have been shortchanged the past few years. Pande has had to have known
that GPUs were under-awarded based on the complexity of the WUs and the quick TPF.
It's a fact a GPU can out compute a CPU. When an optimized core is released that can harness
the power of a GPU,you'll see numbers that will make a 4p blush.
 
There's more to a WU than just atom count, there's also steps involved. With the initial bonus of 10x,
I'd say that GPU folders have been shortchanged the past few years. Pande has had to have known
that GPUs were under-awarded based on the complexity of the WUs and the quick TPF.
It's a fact a GPU can out compute a CPU. When an optimized core is released that can harness
the power of a GPU,you'll see numbers that will make a 4p blush.

Pure conjecture

In the past, GPUs have been able to process only a limited type of Wu whereas CPUs have been Mich more univerrsally useful, therefore CPUs enjoyed a premium on points values.

If GPUs can now handle the same Wu as CPUs, that's one thing. Giving GPUs exclusive WUs does not make sense.
 
PG drove me out of the folding biz last summer. They really don't care about their donors. All they have to do is come out and say point blank that the new GPU wu are the most important wu for their science and all would be well. Unfortunately they can't be upfront about it.
 
There should be two sides to running f@h, the important one that takes care of the science, and the fun and games team who keeps the PPD competitive. Pande says "we have a new WU to run" and gives it to the points team. Points team benches it and sets points to benefit new hardware, maintain competition, and incentivize investment in a particular type of hardware.

Exactly. That's what I thought the DAB was for. A points change such as this should be approved by impartial group of donors before it goes live. Let pg stick to the science, let us decide (through the DAB) how the points should be allocated.
 
Exactly. That's what I thought the DAB was for. A points change such as this should be approved by impartial group of donors before it goes live. Let pg stick to the science, let us decide (through the DAB) how the points should be allocated.

The DAB isn't impartial.

Most conversations devolve into team politics.
 
We also care about points that's why we spend thousands of dollars to get the best ratio points/time$ setup.:D If 3 gtx570 get the same ppd as 4p 6174 for much less $, I will go that route.
 
All the ones I have received have been through v7, on Win 7. I'm no longer running 6.34 anywhere.
It appears they are shuffling these through "in some way". (Just an observation.)
Every 8057 I receive is followed by an 8018 and then back again.
At least for 3 cycles or so on three seperate GPUs.
Also, the 8057s are a lighter load than the 8018s, at least from the standpoint all my GPU temps go up about 3 degrees with 8018.
 
I think the more important question is why am I a member of a 90s death metal band?

In fairness...I think the F@H project being run by a death metal band with Kendrak in it...actually explains a great deal of the questionable decisions made re:F@H
 
I wasn't certain what DAB stood for, so I went to Google....

Now pardon a noobish question on my part...But why is the 90s death metal band Delayed Action Bomb in charge of F@H?


Lol, sounds about right given the antics that go on sometimes.

I cannot agree with teams deciding where the points go. That would simply devolve into teams arguing for the points to go where it suits that particular team.

Points should be allocated by work completed. If it turns out that GPUs are that much more efficient than CPUs, then that should reflect the points that GPUs receive.

What IS important, is that we as donors know that the points are 'correct'. We need to know how PG has determined that 'equal work' is being completed.
Now, at the moment people are up in arms because of a lack of information about the testing methology.
This is a downside of having an open beta forum. People starting 'discussing' (read heated argument) about changes before all the testing and information is known.

As with last time a big change was made (BA8 - BA16) there was lots of debate beforehand based on rumor and not much else.

I think all we can do is put hardware purchases on hold until we know what direction Pande is heading.

When the testing is done and the results are out, then we can make decisions.
All any of the debate is doing is raising ire and blood pressure without possibility of a resolution at this stage.

Everyone should take a deep breath, wait a week or two and see how things pan out.
 
Talk this talk about 'Blah Blah' being more efficient than 'Blah Blah Blah' or 'equal pay for equal play' is all complete bullshit. GPU handle certain WU better because the WU is optimized for that platform and the same goes for a SMP box. There is no way a GPU could process a bigadv WU and a SMP would have a very hard time handling a GPU WU.

Both GPUs and SMPs machines are going to have their share of sweet WUs that really generate points and make their platforms shine. Personally I think it is very poor judgement for PG group to release the new GPU WUs when their is still bad taste in all the SMP folders over the 810x WUs. The PG group could have easily thrown the SMP folders a bone, released the new GPU WUs, and there would not be any of this talk. Then again the PG has never been known for their wisdom.
 
I think the more important question is why am I a member of a 90s death metal band?

Actually the most important question is why aren't I in a 90s death metal band?! Sounds awesome! Beer and tamales every-freakin'-day! <petergriffin> Sweet. </petergriffin>
 
Well, here is the thing, in SINGLE precision floating point, GPU's are quite a bit faster than CPU's. However in DOUBLE precision FP, that is not necessarily the case. A 4P system probably actually has quite a bit MORE DP FP power than a current high end GPU. I dont know the exact numbers off the top of my head. I am not sure if F@H requires DP or if SP is good enough...
 
I think the more important question is why am I a member of a 90s death metal band?


Shhh....

Don't tell Kendrak, but he is not really a member of the band. He is allowed to wear his cheerleader outfit to band rehearsals, and he usually has his pom poms. He thinks they are musical instruments. He is what is known as a groupie.



yougotservedKendrak.jpg
 
Back
Top