Seeking The Best Performance per Watt for Folding@Home

[Olle P]

Originally Posted by Gabriel Torres

In your original message your clearly said VANILLA MULTIMETER. A USD 100 tool isn't a vanilla multimeter.

Well, to me it is, since that's the type I always encounter.
There are those that are cheaper, but they are discarded due to lack of reliability. Then there are those that cost >$300 and they are typically too expensive.

Originally Posted by Gabriel Torres

With a digital wattmeter you don't need to do that -- and it is way cheaper than a Fluke -- and that is why we recommend it over a expensive multimeter.

Wattmeters that cost ~$20 are well known to be unreliable when measuring low power levels, like devices on standby. They're supposed to be extremely unreliable when used to measure the consumption of switched PSUs...

Not really much use to take measures if the results are about as reliable as your best guess, is there?

/Olle

[Olle P]

Originally Posted by Gabriel Torres

It is explained on the tables. "SMP" = SMP client only, "GPU" = GPU client only, "GPU + SMP" = both clients.

For the first batch of tests, yes, but I refer to the 2nd batch, presented on page 5. It seems to imply that the tests were done with the GPU client only, but doesn't state so.

Originally Posted by Gabriel Torres

When I ran both clients on a dual-core CPU the load was at 100% and they were using both cores. Please clarify where you read otherwise.

In the middle of page 3:
"with our Core 2 Duo when we started the GPU client the CPU load went straight to 50%, with one core being idle and the second core being 100% used. ... when we put the SMP and GPU clients to run at the same time consumption (and performance) lowered, as both clients were disputing the CPU."

To me this implies:
- GPU only => One core at 100%, the other one idle. => CPU load = 50%
- GPU+SMP => CPU load < 50% => Both clients try to use that same core, leaving the other one idle.

Cheers
Olle

[vrodic]

Hello guys,

Some time has passed from publishing this article.

I'm curious to know if something has changes, had the performance with new versions of ATI drivers or a new version of folding@home GPU client improved for ATI?

[Cheetos]

Originally Posted by vrodic

Hello guys,

Some time has passed from publishing this article.

I'm curious to know if something has changes, had the performance with new versions of ATI drivers or a new version of folding@home GPU client improved for ATI?

Unfortunatelly not yet.

I don't know if it's a driver or a client issue, I guess that both are the cause of this low performance.

But, the folding guys says that this performance gap is relative to the protein size, in big WUs the gap is around 18%.

Who knows how this issue will be after the DX 11 cards.. I'm waiting.

[vrodic]

Originally Posted by Cheetos

Unfortunatelly not yet.

I don't know if it's a driver or a client issue, I guess that both are the cause of this low performance.

But, the folding guys says that this performance gap is relative to the protein size, in big WUs the gap is around 18%.

Who knows how this issue will be after the DX 11 cards.. I'm waiting.

Hi I've found out the technical reason for ATI being slower on the foldingforum.org. Let's hope this is fixed on 5xxx series.

EDIT: http://foldingforum.org/viewtopic.php?f=51&t=10442

"The primary architectural difference seems to be that Nvidia can store intermediate results in fast short-term memory (like a cache, but managed by the program rather than the hardware) while ATI cannot. For the ATI implementation it is quicker to repeat the calculations than to store to and retrieve from the GPU main memory."

[Cheetos]

Originally Posted by vrodic

Hi I've found out the technical reason for ATI being slower on the foldingforum.org. Let's hope this is fixed on 5xxx series.

EDIT: http://foldingforum.org/viewtopic.php?f=51&t=10442

"The primary architectural difference seems to be that Nvidia can store intermediate results in fast short-term memory (like a cache, but managed by the program rather than the hardware) while ATI cannot. For the ATI implementation it is quicker to repeat the calculations than to store to and retrieve from the GPU main memory."

A huge thanks from me and my buddies !

But there are many arguments there, I guess that more than one are correct since GROMACS is a very "flexible" software.

Who nows if GROMACS 4, recently launched and with GPGPU support, will answer this question. ^^

[Gabriel Torres]

I'd like to add that the main innovation with DirectX 11 is the addition of GPGPU support, providing a well-known programing interface (DirectX) to run "regular" programs on the GPU. In theory after Folding @ Home clients are released supporting DirectX 11 this problem should disapear for the new DirectX 11-based video cards from AMD/ATI.

Cheers,
Gabriel.