Painful days ahead for NVIDIA GPU

[Xilikon]

Now that DanEnsign made this public with this announcement : http://foldingforum.org/viewtopic.php?f=52&t=5452, I wanted to warn everyone that you will painfully see a PPD drop and increased GPU/CPU usage with the new Lambda WU.

new public NV projects 5508-5513, 5017-5022
Hi,

I'm warning you, you will start to see a new series of projects coming out this morning. The proteins we're studying here are larger and in some sense more relevant to biology than anything we've run on GPU2 to date.

However, due to the larger size of the protein, you may see a PPD drop putting NV cards more in line with ATI production. This isn't due to different benchmarking or a different way to assign points, but rather we anticipate NV to be slightly less efficient with larger proteins than it has been with the smaller ones.

Dan

Just to show how painful, I tested them with my 5 GPU during a few days and the PPD drop is close to 10k PPD

 

[crawford60]

ouch. big drop.

oh well, I'm happy as long as I keep crunching out the WUs.

 

[Xilikon]

Yeah, with those WU, NVIDIA and ATI is now roughly the same (ATI gain some and NVIDIA lose).

 

[capreppy]

Will cards with more memory (512 or 1GB) fare better? or is the larger number of shaders on the ATI 4 series cards better for these new WUs?

 

[Wheresatom]

As disappointing as this sounds, those of you that got the ATI cards hoping for a restructuring/optimization are about to get your dividends. I just hope my card can do the most for science possible. Up to that point the points don't even matter. These larger work units might keep the servers from getting to bogged though, so good news there right?!

 

[Xilikon]

Will cards with more memory (512 or 1GB) fare better? or is the larger number of shaders on the ATI 4 series cards better for these new WUs?

Nope, it will still get bogged. The limiting factor is the shaders amount.

My GTX 260 which average 6800-7400 ppd on Supervillin get only 5000-5200 with the new Lambda ones. 8800GS = 2700 PPD, 8800GT = 3100 PPD.

 

[Vediovus]

Woo? Doesn't surprise me that this was coming.

 

[Kendrak]

I'm glad my SMP rigs have been pushing higher PPD with the A2 core.

I might make a serious look into 4750s going forward.

Fold on!

 

[Wheresatom]

I'm glad my SMP rigs have been pushing higher PPD with the A2 core.

I might make a serious look into 4750s going forward.

Fold on!

You bring up a good point. I need to figure out why I am not getting the A2 core SMP units.

 

[metallicafan]

Yeah thats about right. My 8800GS was doing 4K PPD until this morning. Now its seeing about 2600 PPD. If this is indeed better for science than im all for it but I wish they would keep the points a little more consistent.

 

[Kendrak]

Yeah thats about right. My 8800GS was doing 4K PPD until this morning. Now its seeing about 2600 PPD. If this is indeed better for science than im all for it but I wish they would keep the points a little more consistent.

This is one point where I will stick up for stanford, the point are consistant if you are running an ATI card.

They have always said that the benchmark system is an ATI system.

 

[Xilikon]

That's correct... At least, the A2 units seems to compensate a bit for the losses so it's not too painful in the end if you don't run winSMP. To get A2 units, you must run linux or OSX SMP client, not windows.

 

[cereal_killer24]

Why wouldn't they release A2 units for windows smp clients ?

 

[Xilikon]

Why wouldn't they release A2 units for windows smp clients ?

It's because it have lots and lots of issues running under Windows so they cannot release it yet. When they will work out those issues, it will be available so be patient.

 

[metallicafan]

This is one point where I will stick up for stanford, the point are consistant if you are running an ATI card.

They have always said that the benchmark system is an ATI system.

Ah . . I see. I guess I hadnt thought of that. I do run an 3800 as well and that one has been very consistent. I have even seen some increases in that over the past few weeks, so that's nice!

 

[Wheresatom]

That's correct... At least, the A2 units seems to compensate a bit for the losses so it's not too painful in the end if you don't run winSMP. To get A2 units, you must run linux or OSX SMP client, not windows.

I guess I just figured out why I haven't gotten any A2. I haven't really been trying, but I had been paying a bit more attention to what units I was getting. As always thanks Xil

 

[Tigerbiten]

On my farm the drops going to be around 25k PpD ...........

But I still get a better PpDpWatt figure running 2x VM's & 2x GPU's on a single MoBo than just running 2xVM's.
But the difference is a lot smaller.

Luck .............

 

[Kendrak]

Best I can tell is that my Nvidia GPU clients have taken a 40% chop off the top.

This makes getting that 9800GX2 in the mail tomorrow not as sweet

On a side note:

Who is going to get hurt more...... us or OCAU?

 

[Xilikon]

Best I can tell is that my Nvidia GPU clients have taken a 40% chop off the top.

This makes getting that 9800GX2 in the mail tomorrow not as sweet

On a side note:

Who is going to get hurt more...... us or OCAU?

Depends on how much GPU involved. I bet it's more OCAU since a big bunch of our points is still from those who run uniprocessor clients or SMP. My farm is still half SMP so I'm not too much affected but it's still a bit painful to see this but I knew this is bound to happen.

 

[capreppy]

if someone was planning to get the 9800GX2, what would you get to replace it from an ATI perspective to maintain the same level of points? Will the 4850/4870 keep up with the 9800GX2? Will the 4870x2 keep up? I know the 4870x2 is $500 per, but what kind of output can one expect from it?

Or is the 9800GX2 (4 GPU in 2 PCIe slots) still effective, but it won't product the 20K points previously. Maybe 16K points?

 

[jws2346]

That's correct... At least, the A2 units seems to compensate a bit for the losses so it's not too painful in the end if you don't run winSMP. To get A2 units, you must run linux or OSX SMP client, not windows.

Yeah bro Xilikon, that's what's kinda' scary, no A2 clients on Windose and no GPU clients on Linux (Hmm...., where oh where is Stanford, I ain't about to hold my breath until they come up with a A2 client for Windose or a GPU2 client for Linux, d'oh...)

I've got a little 40 gig HD comin' by ebay jet or whatever and when it gets here I'm gonna shut down my Windose quad with the new 8800gt and play "musical hard drives". I'll put the 40 GB HD in the quad, load Ubuntu 64bit v8.04 on it, try to follow the directions by Shellnut2 et al, and see for my own curiosity if it's actually true you can use a "CUDA" capable GPU card to fold under Linux. (and in the meantime "crunch" the hell outa' A2 jobber doobers ) Or........, If That's a gimongous if, the things I've read are true that there is a program out there already that DC's on a "CUDA" nVIDIA 88xx series GPU card under Linux I'll look into that. (I think it has something to do with WCG ) If I had anything to do with F@H I'd be a lttle concerned and get offa' my "tush" and maybe speeds things up a little ("tush" = PC for a$$)

Oh well, I'm just gonna' do the best I can with what little bit I have.

DCin' for the CURE

 

[Tigerbiten]

.........
Or is the 9800GX2 (4 GPU in 2 PCIe slots) still effective, but it won't product the 20K points previously. Maybe 16K points?

A bit less.
My 9800GX2's have dropped from 5,000 PpD to 3,800 PpD.
Amd I've upped the shaders by another 108 Mhz to 1783 Mhz get that output.
So what was a 40k box is a 30k box.

Luck ...........

 

[Jathanis]

Having just gotten my first TWO DAYS of use out of my 9600GSO, now I get to take a loss... Great news... Well, it's been a great two days, at least!

So, from the experts, since I'm running the afore mentioned GPU and an OC'ed e2160 dual core, am I better off keeping my rig the way it is now, GPU2 & WinSMP clients, or trying to run a VM setup for SMP with the GPU? Or can a dual core chip even support thw VM setup if the computer is to still be used for other things? Or should I just forget the CPU folding and move that to WCG? I'm so confused! (Not that that's unusual...)

 

[Sunin]

4870x2 would be nice capreppy!

And remember it effects everyone so yeah you drop some, but so does the guy your about to run over... should overall be a pretty whatever... event...

 

[MrGuvernment]

Woo? Doesn't surprise me that this was coming.

it shouldn't since standford all along said NVIDIA would get a cut back of PPD.

 

[drgnfang]

Mine are all still chugging along with 5506 (and one 5013). Points are actually up.

Here is the question of the hour: Why aren't the larger number of shaders in the gtx280 cards keeping those cards from going down? Are they still only using like 30% of them, even with the larger WU?

'cause that would be pretty sad.

 

[evilr00t]

8800GTS 320M 621/1512/792 went from 4500PPD (4489) to 2600PPD (2580). Seems rather steep to me... also, GPU temp went up 3 degrees from 63 to 66C. Me not like!

 

[drgnfang]

Temps going up is bad. My office already hits the mid-90's when I'm not here during the afternoon.

 

[BillR]

no comment,,,,,,,,,,,,

 

[Pooch666]

Holy Shit! I got the new client a while ago and went from 1.26 to 2.08. Whatta bite in the ass

 

[Arcygenical]

Well then, what's the new king for PPD/$...

I mean, the 8800GS seems so much less attractive now... Might still be on top though.

 

[D3v01D]

Personally, It doesn't matter to me (the ppd thing).

and in some sense more relevant to biology than anything we've run on GPU2 to date.

This is what matters to me.

 

[Tigerbiten]

A mixed folding farm of a Q6600 running 2x VM's with 2 or 3 cards running the GPU client is looking better all the time, now that the nVidia client has dropped in value but the SMP client has increased in value with the new A2 core.
With that type of setup you should still get above 35 PpDpW.
Plus the SMP work-units take longer to crunch when Stanford goes down again ............

Luck .............

 

[MrBooMY]

It annoys me a lot that not only do the new blocks take LONGER to finish, but are worth LESS points. That seems kinda messed up to me. It's still for a good cause, and I'm not going to stop, but that logic doesn't make sense to me at all. Takes LONGER, LESS points. It seems like it should be, Takes LONGER, MORE points. My 8800 GTS 640MB went from 4600 PPD on most blocks down to 2580PPD. Not only that but now my card went from 73-75C to now 77-79C with the new blocks as well. More heat = more electricity! Standford can bite my balls on this one, they screwed us.

 

[HighYield]

Personally, It doesn't matter to me (the ppd thing).

"and in some sense more relevant to biology than anything we've run on GPU2 to date."

This is what matters to me.

+1
I would rather have them working on something more useful. As Sunin was pointing out it just won't matter if everyone's production takes a dive.

 

[MrGuvernment]

It annoys me a lot that not only do the new blocks take LONGER to finish, but are worth LESS points. That seems kinda messed up to me. It's still for a good cause, and I'm not going to stop, but that logic doesn't make sense to me at all. Takes LONGER, LESS points. It seems like it should be, Takes LONGER, MORE points. My 8800 GTS 640MB went from 4600 PPD on most blocks down to 2580PPD. Not only that but now my card went from 73-75C to now 77-79C with the new blocks as well. More heat = more electricity! Standford can bite my balls on this one, they screwed us.

NO they didn't - ALL along they said the NVIDIA ones would get a cut in PPD just wasn't sure when, that when is now, so stop crying that something they said would happen DID happen, you dont like it, dont fold, it isnt for the point anyways and if it is your in it for the wrong reasons. It is not like this only affects you, everyone will drop, so nothing changes but you get less points, so it causes more heat. as they said go back to the old client if you want the old point scale.

 

[crawford60]

It annoys me a lot that not only do the new blocks take LONGER to finish, but are worth LESS points. That seems kinda messed up to me. It's still for a good cause, and I'm not going to stop, but that logic doesn't make sense to me at all. Takes LONGER, LESS points. It seems like it should be, Takes LONGER, MORE points. My 8800 GTS 640MB went from 4600 PPD on most blocks down to 2580PPD. Not only that but now my card went from 73-75C to now 77-79C with the new blocks as well. More heat = more electricity! Standford can bite my balls on this one, they screwed us.

are you seriously that annoyed?

Its just some points. get over it. the goal is to do calculations in hope of understanding and curing some diseases. If anything you should brag about WUs, not points. Points are just fun extras.

 

[chrisf6969]

I just got a 5508 protein. It took 1hr 20mins, instead of the usual approx. 1 hour, 5 mins. (with a GTX 260). AND it was only worth 430 points!!

So it takes longer and is worth less points!!

To be exact 23% longer for 10% less points.

Or a total of 37% less PPD according to my calculations.

480PPU @ 65 mins/unit vs 430 PPU @ 80mins/Unit

 

[drgnfang]

Yeah, my 88gt's just got some of the new WU. 5k - 5.3k PPD down to 3.3k - 3.4k PPD.

While the lowering of the PPD in general doesn't bother me (affects everyone, blah blah blah). What does bother me is the fact that they are once again saying "This is more better science" while devaluing the WU with lower points. I would have preferred to see the the credit for the older (less science) WU drop and keep the newer (more science) WU at the same credit, or slightly higher since it takes longer to crunch.

Lets be honest here, as much as we are all here for the science, if it wasn't for the points and competition the [H]orde wouldn't exist. Most of us would still be running a client here or there on our spare hardware, but you would never see things like people running 4x 280gtx's or 4x98gx2, or several hundred single client boxen (we miss you FLECOM!).

Ah well, at least we can say it is never boring here in F@H land.

 

[d00dz]

drgnfang: Yea I agree with you there.

Running 10 GPU clients each was getting 5-5.5k ppd now with the actual real wu work on science we are getting less ppd. Doesnt give much incentive for people to build more boxes.