ATI multi GPU

[thilanliyan]

http://www.insidehw.com/Editor...lution-for-masses.html

This is interesting:
?Basic mistake is to expect higher performance levels with single frame buffer for both GPUs. That can be applicable to system memory (RAM) that has few applications that have access to it. When memory on graphics cards is concerned it?s not the case and gain in performances wouldn?t be great. Off course complexity of implementing that solution shouldn?t be disregarded along with expenses that would follow that kind of development and it is actually cheaper to implement double amount of video memory in this case?.

 

[apoppin]

keep on reading

Good example of faster hardware then software development is Side Port technology. This technology supports direct link between two memory buffers on graphics cards that allows speeds in communication up to 21.8GB/s as opposed to PCI express link with maximal bandwidth of 8GB/s. At this moment even PCI express link is enough since maximal bandwidth used for communication between two frame buffers is less than 6GB/s because there is no software that can use more than that.
Since recommended price for 4850x2 is around 399$ it is clear that instead of priciest GDDR5 cheaper GDDR3 will be used. Both x2 cards from 48 series have 2x1GB of video memory. We were curious what effect on price and performances would have a graphics card with less memory (2x512Mb). By Mr. Marinkovic?s opinion it all depends on what game (benchmark) you are using. At the moment 2x512MB would be enough but latter with games that could take the most of larger frame buffer we could have large impact on performances.

Our experience with Crysis benchmark confirms such opinion. With Radeon 3850 with 256MB and 512MB memory we saw big differences in performance levels although we didn?t notice that kind of difference in other games

i guess i will find out what sort of framebuffer limitations i will see when i stick X3 back in tonight and limit the 2GB to a single FB of 512MB

 

[thilanliyan]

Originally posted by: apoppin
keep on reading

Good example of faster hardware then software development is Side Port technology. At the moment 2x512MB would be enough but latter with games that could take the most of larger frame buffer we could have large impact on performances.

Our experience with Crysis benchmark confirms such opinion. With Radeon 3850 with 256MB and 512MB memory we saw big differences in performance levels although we didn?t notice that kind of difference in other games

i guess i will find out what sort of framebuffer limitations i will see when i stick X3 back in tonight and limit the 2GB to a single FB of 512MB

No the frame buffer part I quoted was about having a SHARED framebuffer between both GPUs...which currently doesn't exist and which Derek Wilson actually said is how it should be I think in his blog. I was just quoting it because the person being interviewed said it's not the "holy grail" it's made out to be.

 

[betasub]

It may not turn out to be such a "holy grail", but it sure would be nice to make better use of a scarce supply of expensive vRAM (such as GDDR5).

 

[apoppin]

Originally posted by: thilan29

Originally posted by: apoppin
keep on reading

Good example of faster hardware then software development is Side Port technology. At the moment 2x512MB would be enough but latter with games that could take the most of larger frame buffer we could have large impact on performances.

Our experience with Crysis benchmark confirms such opinion. With Radeon 3850 with 256MB and 512MB memory we saw big differences in performance levels although we didn?t notice that kind of difference in other games

i guess i will find out what sort of framebuffer limitations i will see when i stick X3 back in tonight and limit the 2GB to a single FB of 512MB

No the frame buffer part I quoted was about having a SHARED framebuffer between both GPUs...which currently doesn't exist and which Derek Wilson actually said is how it should be I think in his blog. I was just quoting it because the person being interviewed said it's not the "holy grail" it's made out to be.

Well the X2 GPUs are still *sharing* a [much larger] Framebuffer, aren't they?
- since each GPU has 1GB each of vRAM

i have CF X3 running right now .. and my frame buffer is now more limited by using a 512MB card with 2 1GB 4870 cores, isn't it? - to 512MB *shared* .. down from 1GB shared by the x2

NOW this IS interesting .. Running the DX10 benchmark PT Boats, Knights of the Seas, i get excellent Frame rates with 4870x2 and 280GTX [40s and 50s]
-using a single 512MB 4870, the frame rates drop - from ~50fps to ~5fps

NOW ,,, *adding* a 512MB 4870 to a 4870x2 for CFx3, drops the frame rates [from 50 fps] to slightly better than a single 512MB 4870 [~5 fps]
- framebuffer?

what do you think?

anyone ELSE want to try this benchmark?

http://www.gamershell.com/download_21133.shtml

especially if you have a 1GB or 512MB GPU .. post your results

 

[thilanliyan]

Originally posted by: apoppin

Originally posted by: thilan29
No the frame buffer part I quoted was about having a SHARED framebuffer between both GPUs...which currently doesn't exist and which Derek Wilson actually said is how it should be I think in his blog. I was just quoting it because the person being interviewed said it's not the "holy grail" it's made out to be.

Well the X2 GPUs are still *sharing* a [much larger] Framebuffer, aren't they?
- since each GPU has 1GB each of vRAM

The shared framebuffer is different to what you're describing...it would be accessible to both cores and would allow the card to act as if it didn't have 2 separate GPUs thereby ridding us of scaling issues...at least that's the theory of it as I understand.

 

[VirtualLarry]

Originally posted by: thilan29
No the frame buffer part I quoted was about having a SHARED framebuffer between both GPUs...which currently doesn't exist and which Derek Wilson actually said is how it should be I think in his blog. I was just quoting it because the person being interviewed said it's not the "holy grail" it's made out to be.

A shared framebuffer would halve memory bandwidth going to each GPU, neccessitating even more expensive DRAM to achieve the same performance that you could get by using seperate DRAM banks for each GPU. Not to mention, it would cost significant engineering expertise to pull it off.

Not worth it.

 

[VirtualLarry]

Originally posted by: thilan29
The shared framebuffer is different to what you're describing...it would be accessible to both cores and would allow the card to act as if it didn't have 2 separate GPUs thereby ridding us of scaling issues...at least that's the theory of it as I understand.

It wouldn't get rid of scaling issues at all.

 

[thilanliyan]

Originally posted by: VirtualLarry

Originally posted by: thilan29
The shared framebuffer is different to what you're describing...it would be accessible to both cores and would allow the card to act as if it didn't have 2 separate GPUs thereby ridding us of scaling issues...at least that's the theory of it as I understand.

It wouldn't get rid of scaling issues at all.

As I said that's how I understood it (that the OS would only see one card and the driver would also see it as one card (ie. one massive 4870x2 with 1600 shaders and ONE 2GB framebuffer))...here's a quote from Derek Wilson's blog:

"We hope, if single-card multi-GPU solutions will be the continuing focus of AMD, that we'll see fundamentally better multi-chip architectures down the line (once again, we need a shared framebuffer). "

Why is it that he says we need a shared framebuffer?

 

[BFG10K]

A shared framebuffer would solve the issue of data duplication but it would not solve the issue of scaling. As long as the GPUs are separate entities you're going to have scaling issues.

It's kind of like how a shared cache on a Core 2 duo doesn't magically make all applications run twice as fast. You still have to program the software to take advantage of two CPUs.

 

[SSChevy2001]

A shared frame-buffer isn't going to solve the problem. AFR is really the problem, and IMO Lucid's Hydra chip seems to be a better route for multi-gpu setups in the future. Just like how the workload will be broken up for each GPU instead of this whole clone the memory between GPUs.

 

[thilanliyan]

Originally posted by: BFG10K
A shared framebuffer would solve the issue of data duplication but it would not solve the issue of scaling. As long as the GPUs are separate entities you're going to have scaling issues.

It's kind of like how a shared cache on a Core 2 duo doesn't magically make all applications run twice as fast. You still have to program the software to take advantage of two CPUs.

Ah okay thanks my understanding of what was possible was way off. Is it possible to make it seem as if say 2 4870s with 800sp each act like one monster GPU with 1600sp other than actually making a single chip that size or is that a pipe dream?

 

[BFG10K]

Originally posted by: SSChevy2001

and IMO Lucid's Hydra chip seems to be a better route for multi-gpu setups in the future. Just like how the workload will be broken up for each GPU instead of this whole clone the memory between GPUs.

Hydra looks good in theory but I'd wait for actual results before passing judgment. Intercepting API and/or driver calls like that has the potential to create an absolute clusterfuck of problems, maybe even more than AFR.

Originally posted by: thilan29

Is it possible to make it seem as if say 2 4870s with 800sp each act like one monster GPU with 1600sp other than actually making a single chip that size or is that a pipe dream?

I don't believe so. In order to attain the level of integration required for multiple cores to truly behave as one (i.e. all hardware resources completely shared and visible to each other), you'd basically come full circle to having one core.

It would be very expensive and difficult to implement the level of sharing required in order for multiple cores to gain what comes naturally with one core.

Again think about multi-core CPUs. What would it take to truly combine the execution resources of a quad-core to make it run applications four times as fast as the same clock single core? You?d basically have to integrate the four cores into one giant core.

 

[Dadofamunky]

Wow. Sounds like X3 scaling simply doesn't happen - in fact, it ANTI-scales. X3 just does not look like a good option at all!

But X4 works fine, right? At least you wouldn't go from 50 fps to 5. Or lose that huge chunk of RAM!

But you still won't get huge scaling in rates, based on the reviews I've seen, beyond the first X2. Seems like the marginal utility of buying a second 4870X2 card is probably somewhat limited. What, like a 20% average increase from the second card?

Those are the benchmarks I'D really like to see.

It seems evident that ATI did not design as thoroughly or effectively for odd-number scaling. This makes sense, in computer math. with binary words and longwords being what they are (0,1,2,4,8,16,32,64,...) No odd numbers in the binary continuum. Or, they marketed a feature that did not have enough QA cycles or sustaining engineering support. Or the hardware team simply left or was laid off.

This makes me wonder - how effectively would Hybrid Crossfire work? Based on this background, I would tend to doubt its effectiveness.

 

[apoppin]

x3 works .. for some games .. for others it is a *disaster* .. but still much improved over the clusterfsck that was Cat 8.8 and X3

i have ALL my 4870-512M/4870x2 and 4870+4870x2 = x3 benches done with Cat 8.9 !!
[also 2900xt/8800gtx/280GTX benches with their latest drivers also]
i am rebuilding my PC today around x48 MB and will repeat all my benches with it

what do you want to see? .. i am tearing my PC down shortly

i'd still love to see anyone else's benchmark results with either a 512MB GPU and a 1GB GPU

http://www.gamershell.com/download_21133.shtml

it is primarily a DX10 benchmark .. my 512MB cards run it like crap [fully maxed at 19x12 with 4xMSAA] and my 1GB cards run it great

 

[Dadofamunky]

Originally posted by: apoppin
x3 works .. for some games .. for others it is a *disaster* .. but still much improved over the clusterfsck that was Cat 8.8 and X3

i have ALL my 4870-512M/4870x2 and 4870+4870x2 = x3 benches done with Cat 8.9 !!
[also 2900xt/8800gtx/280GTX benches with their latest drivers also]
i am rebuilding my PC today around x48 MB and will repeat all my benches with it

what do you want to see? .. i am tearing my PC down shortly

i'd still love to see anyone else's benchmark results with either a 512MB GPU and a 1GB GPU

http://www.gamershell.com/download_21133.shtml

it is primarily a DX10 benchmark .. my 512MB cards run it like crap [fully maxed at 19x12 with 4xMSAA] and my 1GB cards run it great

Dual X2 rig if possible - I know I'll never run it, but I'd like to see some real-world performance numbers...

 

[apoppin]

Which ones? i finished all my benches with Cat 8.8 and 8.9 on my P35MB

i have 4870/512 + 4870x2 =x3 CF. i just installed my X48 MB yesterday. .. that takes some time just to set up the 10 games and all my tools and aps.
. . . also i have 4870x2 and 4870 benched separately
.. with 280GTX tossed it for a reference .. but remember this is on a 1.0 PCIe board with 16x + 4x PCIe for CF. My x48 is 2.0 with a full 16x+16x PCIe CF. All of the benches i previously ran on P35, i am repeating for a MB platform comparison to x48 with my e8600 still @ 4 Ghz.

10 games ... Fear, Hl2, ETQW, UT3, Crysis, LP, CoJ, STALKER & GRID .. inc. PT Boats which isn't yet out.

 

[shangshang]

multi-gpu has been an experiment and still is. I guess I don't understand why so many people are willing to shell out hard earned money just to become ginuea pigs. I'll give multigpu solution another 2 years before i'll consider.

 

[apoppin]

Originally posted by: shangshang
multi-gpu has been an experiment and still is. I guess I don't understand why so many people are willing to shell out hard earned money just to become ginuea pigs. I'll give multigpu solution another 2 years before i'll consider.

no it is no experiment

MultiGPU works!

The X2 clearly beats the 280GTX, for example, and it is a generally a pleasure to run