It seems as though the combined OpenCL tests just divide the workload up evenly amongst all GPUs present.  3 GPUs = 33% each while 4 GPUs = 25% workload each.

Factor the Qual-SLI score by multiplying it by 4, and that would be the "theoretical" full 100% utilization score.  2549 x 4 = C10196

On the ATI side, it only uses a single GPU regarless of how many GPUs are installed on the system, BUT it uses almost 100% utilization of that single GPU.  Take my C4400 score and multiply that by a factor of 4 to get what 4 5870s would "theoretically" score (4400x4=C17600)

So, 4 GT200 (240SP) GPUs would theoretically score a C10196 if all GPUs were utilized at 100%.
And, 4 Cypress (1600SP) GPUs would theoretically score a C17600 if all GPUs were utilized at 100%.

So, we can conclude that RV870 (1600SP) is 73% faster than GT200 (240SP) in doing OpenCL calculations.

Now, we have to remember that both company's OpenCL drivers are in their infancy as well as the API itself.  From this, we can conclude that improved OpenCL drivers could change the performance drastically in future driver releases.  If both companies improve their drivers at the same rate, we can see that the ~73% figure may remain in effect for quite some time if not indefinitely.

Let's see what Fermi brings to the table.  Let's see if GT300 is 73% faster than GT200 at OpenCL as well.

Talonman

I would be on-board with that, except for post 11 in this thread...

http://forums.evga.com/tm.aspx?m=86719

I still question if this score = actual performance in OpenCL apps, when comparing ATI to Nvidia..

I posted a reply.  Also, you have to realize that those are "OpenGL + OpenCL" hybrid demos so they are not purely OpenCL as the DirectCompute benchmark tests.  Also, those tests are not indicative of OpenCL's GPGPU performance.

We did have had a 5850 score reported...
http://www.ngohq.com/grap...mark-10.html#post87400

So that means:
A 8800 GTS and a single 4850 produces around C453

A single 260 produces around C707

A single XFX HD 5770 1GB produces around C1042

A single 295 produces around C1431 using both sides of the GPU...

A single 4890 produces around C2350

Two 295's @stock running Quad-SLI score around C2549

A single 295 and single 280 OC'ed produce around C2575

A single HD 5850 @ stock produces about C3034

Luv's single 5870 produces around C4058 - C4405

And still trying to figure out why the scores are so high on ATI GPU's...

Talonman;87409Thanks for the info...  :)

The thing that bothers me is 1/2 of my 295 produces more FPS, than Luv's 5870 in this app...
http://forums.evga.com/tm.aspx?m=86719

Examples on my system:




If he actually can run OpenCL apps faster on his single 5870, than 1/2 of my 295, why can I generate higher FPS in other OpenCL apps?

I guess a 5770 is actually 1/3 faster than a 260 too... NOT!!
http://www.techspot.com/r...eon-hd-5770/page4.html

Or how a 4890 is almost 2X as fast as 1/2 of a 295. (GTX 275)

Hopefully Pat will get back to me soon with his answer... 

Not stressing... Just love to understand what is going on.

OpenGL/OpenCL apps is not a good enough answer. That doesn't justify the large discrepancies in scores between vendors.

I am going to ask Pat to install a processing timer. I want to see if a single 5870 actually completes the test, in almost 1/2 the time it takes my 295 and 280 to complete the test.

My money says that might flush the truth out.

Talonman

Not stressing... Just love to understand what is going on.

OpenGL/OpenCL apps is not a good enough answer. That doesn't justify the large discrepancies in scores between vendors.

I am going to ask Pat to install a processing timer. I want to see if a single 5870 actually completes the test, in almost 1/2 the time it takes my 295 and 280 to complete the test.

My money says that might flush the truth out.

This again will probably make your head spin a little bit.  Here is another "single" GPU run of the DirectCompute benchmark 0.45:













I ran it again and got around 4400 so that was just fluke # I guess.  Well, that shows that ATI and/or Pat has some work to do yet

edit:  4400 is what happens when the GPU Core is at 925mhz...  Ooops :)  I guess I could hit 5000 with an OC a tad higher than 1000 "on a single GPU". 

luv2increase

This again will probably make your head spin a little bit.  Here is another "single" GPU run of the DirectCompute benchmark 0.45:
 
I ran it again and got around 4400 so that was just fluke # I guess.  Well, that shows that ATI and/or Pat has some work to do yet

C4058.2  Wow!! 

Nice score no matter what!

Just to add more confusion...

http://www.ngohq.com/grap...mark-10.html#post87400

Talonman;87436FYI - I just timed my 'Combined OpenCL' run...

I am 3 seconds shy from hitting a 2 minute run time. (1minute, 57 sec)

I would love for an ATI guy who can generate a higher score than C2575 to time his run, and post up the processing time it takes to complete the test.

thehippo;874471 minute 13 seconds @ 905mhz (C3792)
1 minute 31 seconds @ stock (C3042.5)

(He is running a single 5850)

Talonman;87449Gasp!

Thanks...   (I think?)   :)

I don't know what to think now...

C4485.6 = 60 Seconds Exactly

I don't know what you're talking about.  I cannot do the "combined OpenCL" test.  There is a known problem with this program that says there is only 1 OpenCL device.  On that forum Pat is on, they discuss it.


That C4485.6 is straight GPU with 0% Utilization on the CPU.  My CPU doesn't do anything on the "OpenCL" test of this benchmark.


edit:  It is obvious that since it "does not" take everyone the same exact amount of time to complete this test as I thought it did until you posted about the guy finishing the test at two different times depending on what his OC was, and his scores refected as much too.

Deducing from this, it is obvious that there is a pre-determined number of problems or problem size that the test sees how fast your GPU can "complete" the task.  The "faster" you "complete" the task, the higher your score is.  It seems the ATI architecture completes the task at a very fast rate while the Nvidia cards, although each card is even utilized, has a hard time processing whatever problem(s) the programmer Pat implemented.

We can gather that this would indeed either be:

a) An Nvidia OpenCL driver issue
b) An Nvidia architecture issue

Talonman

freakysqeeky

Yeah it might be both. Nvidia benches look right..Except for talonman's 295 with the 280 is benching higher than a quad?

Two 295's @stock running Quad-SLI score around C2549

A single 295 and single 280 OC'ed produce around C2575

I think stock and OC is the key there?

Does look odd though!

A single 295 and single 280 OC'ed produce around C2575

But there is no application that would use the 280 (as a physx card) to its fullest. Except for this application
Unless that is about to change in the near future?
Two 295's @stock running Quad-SLI score around C2549

A single 295 and single 280 OC'ed produce around C2575
That does look about right, if it wasn't a physx card
Your losing your grip here talonman :)

Talonman

Well it's in operating dedicated PhysX mode, but it's being used a an OpenCL processor here in this app...

OpenCL doesn't care, it can still allocate work to my 280, providing the programmer is supporting multi-gpu's in his app.

I don't believe it is processing PhysX, nor do I believe this app contains any Nvidia PhysX.

I think GPU's given the name 'Dedicated PhysX Processor' are now used for OpenCL, CUDA, and Ditect Copmpute apps as processors.

He was using (4) stock OpenCL processors, and I was using (3) OC'ed OpenCL processors.

Dedicated PhysX cards can calculate at least 4 things now.
( PhysX, OpenCL, CUDA, and Ditect Copmpute.)
And the reason I think they have a bum name...

The name 'Dedicated PhysX Processor' implies that it can only calculate PhysX.

True, but we both knew what all three cards could & can do if all the resources are being used, especially all three of your cards. Is this benchmark really telling us what we already knew? No big breakthrough, or am i missing something?
"He was using (4) stock OpenCL processors, and I was using (3) OC'ed OpenCL processors."
I understand that its not physx and that's a start for sure.
Not banging on the benchmark its a great start, but it would be nice to see something useful actually using "all" that we have.

My 295 can't take anymore, so the cpu must be taking some of the load, were being bottlenecked? I'm not sure what's going on.
I hope its a driver issue and not An Nvidia architecture issue like luv said. 

these scores make allot of sense, in raw compute power the ati 5000 series is crazy powerfull around 3 tera flops per gpu. however getting any thing close to that in most real world apps is impossible. this bench mark seems to be very compute intensive which is very nice for a synthetic bench mark ... a single g200 core has around 0.9 tera flops of theoretical compute power. Seems there is a problem with the multi gpu scores they make less sense ...

The fermi when ever it comes out should have around the same compute power as the ati cards.