Unified GPU/SMP benchmarking scheme: equal points for equal work

[Ben Lamb]

Check this out !

Does this mean Bigadv has had its day and the return of the (nvidia) GPU farm is coming again ?

Is it time to put the xeons/opterons on ebay and invest in nvidia cards - looks like nvidias help in producing a cuda client is paying well for them.

Or is this all because of advances made in gromacs 4.6

Thing is what do they mean by equal points for work done ?

You have to consider the average GPU work unit contains only 1980 atoms while a bigadv has 1340422 so totaly different. One performed in a gravity field and the other in water with loads of molecules. Does this mean a GPU can fold a bigadv if it has enough memory ?

No one can answer these questions as yet but the Pande group is starting to get on my nerves. We all bought GPUs for folding and then cpus and now it looks like GPUs again. This equal points system should of been used in the first place and propper clients produced for both amd and nvidia cards. I am starting to wonder if the evil lure of hard cash is behind some of this as folding@home has got to be selling a lot of hardware due to all this chopping and changing. I want to contribute to help sick people get better not fill the pockets of the likes of Nvidia or provide free research for rip off drug companies. Starting to smell a bit of a rat here (probably my socks)

If all types of calculation can now be done on the gpu will that make cpus pretty much redundant for folding with the amount of gigaflops a gpu can process compared to cpu ?

This is the proof of what this could mean for cpu folders !

http://ambermd.org/gpus/benchmarks.htm#Benchmarks

Heres the official statement -

current benchmarking calculations for SMP and GPU projects are performed on different machines since originally the SMP cores could not perform the calculations that the GPUs cores could and vice versa (GPUs were only for implicit solvent calculations and SMP only for explicit solvent calculations). With recent advances in both cores and completion of our testing of these capabilities to ensure agreement, we are now confident we can do the same work on both cores. Thus, we feel that it is time to unify GPU and SMP benchmarking, both for simplicity and fairness.

In order to complete the move towards this plan of "equal points for equal work," new GPU projects will be benchmarked using the existing SMP benchmarking scheme. Based on our internal tests, the end effect of this new, unified benchmarking scheme would boost the points for the GPU projects, both in terms of base points but also by bringing Quick Return Bonuses to GPU clients. In order to test the new scheme, we have started a GPU3 project (Project ID: 8057) and released it for beta testing. Once the benchmarking scheme has been tested, all the current GPU projects will be re-benchmarked to reflect the changes in the benchmarking scheme.

[kirk46]

interesting...

looks like GPU
WU's are going to get the same bonus points as SMP WU's do...

i wouldn't mind getting that WU to see what the difference in points are

[cave_diver]

Aren't the CPU units considerably bigger, so surely that means GPU units are worth even less? Given that the only way I've managed to get a modest PPD is because of collecting GPU's, for them to axe the points would hurt, BADLY!

Thankfully I'm not as bitter as the more "pro" folders with huuugely expensive rigs as all the other points cut were, quite literally, way above me. But now this hurts! I depend on my GPU's! then do about 80% of my work!

[Ben Lamb]

Originally Posted by cave_diver

Aren't the CPU units considerably bigger, so surely that means GPU units are worth even less? Given that the only way I've managed to get a modest PPD is because of collecting GPU's, for them to axe the points would hurt, BADLY!

Thankfully I'm not as bitter as the more "pro" folders with huuugely expensive rigs as all the other points cut were, quite literally, way above me. But now this hurts! I depend on my GPU's! then do about 80% of my work!

No dont worry GPU points will be going up considerably but it looks like possibly only for nvidia cards at first so you will be doing a lot better. GPUs and CPUs will be folding the same work units with the new sceme so bigger points for the GPU which will be performing explicit calculations as well. Where this leaves cpus I realy dont know.

CHECK THIS OUT !!!

http://ambermd.org/gpus/benchmarks.htm#Benchmarks


Looks like a gtx680 is orders of magnitude faster than cpus at explicit molecular calculations looks like keplers are the things to have for the future. Should of predicted this realy.

[Chicken76]

What does "2xGTX580 (1 node)" mean?
Could that be SLI?

[Christopher N. Lew]

Well, this should be fun (except for people moaning and tossing wild accusations around).

My reading is that existing SMP WUs will not be changed, so that people with multi-processor machines will not be loosing anything. There will now be a Quick Return Bonus scheme for GPUs, which will surely increase their points. But I'm sure that some WUs will run faster with the large number of cores of a GPU, and others will be faster with the larger memory accessible to CPUs. I hope that Stanford's bench marking scheme will reward we donors accordingly, though history suggests there will be huge numbers of complaints, whatever is done.

Bear in mind that Stanford's advice is not to buy hardware solely for running particular clients that exist today, because they may not exist tomorrow.

And I have no doubt that sometime the client will be revised to run CUDA 5, which will probably mean Kepler GPUs get a big increase over previous generations of cards.



Anyone else pondering an Adapteva machine http://www.adapteva.com/ ? Only 2 days left!

[Ben Lamb]

Originally Posted by Chicken76

What does "2xGTX580 (1 node)" mean?
Could that be SLI?

Sort of - basicaly multi gpu folding is coming you dont need sli just a bunch of cards in one system its been around for a while now for raytracers etc. Future of folding is four gpus shoved in a box running the new work units getting huge ppd !

That is if Nvidia doesnt do the usual thing and somehow limit this type of thing to expensive tesla cards.

But looking at the gromacs 4.6 documentation that future F@H cores will be based on consumer cards are fine but need to be gtx460+ performance wise to make efficient use of the new shared GPU/CPU calculations if these are used by F@H in the future. Its all very interesting but its going to cause a few headaches for folders setting it all up but that is part of the fun !

[Ben Lamb]

Originally Posted by Christopher N. Lew

Well, this should be fun (except for people moaning and tossing wild accusations around).

My reading is that existing SMP WUs will not be changed, so that people with multi-processor machines will not be loosing anything. There will now be a Quick Return Bonus scheme for GPUs, which will surely increase their points. But I'm sure that some WUs will run faster with the large number of cores of a GPU, and others will be faster with the larger memory accessible to CPUs. I hope that Stanford's bench marking scheme will reward we donors accordingly, though history suggests there will be huge numbers of complaints, whatever is done.

Bear in mind that Stanford's advice is not to buy hardware solely for running particular clients that exist today, because they may not exist tomorrow.

And I have no doubt that sometime the client will be revised to run CUDA 5, which will probably mean Kepler GPUs get a big increase over previous generations of cards.



Anyone else pondering an Adapteva machine http://www.adapteva.com/ ? Only 2 days left!

No sign of the huge number of complaints yet I have tried to start the ball rolling but noone seems arsed as yet over at standford. I have been told to use diverse hardware which went down like a lead balloon

As for the kepler increases - you wont see them just 50% faster than a gtx580 for the gtx680 so inline with the shader count and no more.

[cave_diver]

Thanks for the re-assurance Chris!

So really this is good news for GPU folders - the QRB for the GPU has been needed for ages. And it also rebalances the points for the huuuge amounts of power the GPU's are consuming too so now the PPD/W calculations are getting a bit closer!

You don't think nvidia has nudged stanford and said "our sales are down a bit, do something".