R520 info thread #(9!)^9

isnt the theoretical maximum 32%? not that it would work out that way usually as seen in the other benchmark (css?)

 

cant edit, 30.9 not 32

 

It's always nice to quote yourself. I just wanted to say that apparently Ati succeded in creating an architecure which can be clocked quite high. The XT it is not clocked that high just to be competitive. It can do more. And to repeat ... question is the XL. Actually, why didn't they OC the XL too ? If the XT is such good OC-er, there must be some reason for clocking the XL significantly slower. 125 Mhz is quite a difference.

 

They didn't have one, only an XT.

There's some tasty XL OCing info over here:

Doesn't look like the XL is too shabby, needs better cooling though.

 

Core and memory clocks don't add. If you're completely memory bandwidth limited, then the maximum you could get out of the above overclock would be 20%. If completely core limited, 10%. If somewhere inbetween, then you'd get an improvement somewhere inbetween 10% and 20%. If CPU-limited, you'd get less of an improvement.

There's no logical reason why they should be getting a 25% improvement.

 

You know, I really don't understand why people are saying these are good overclockers. They're hot as hell, and while the cores may be capable of high clocks, you have to have incredible cooling to do it. Hell, that one guy reported that his XL was running at 80C just at stock speeds....had to put a waterblock on it just to get it down to 63C, at stock speeds!

And a 10% core overclock really isn't much to write home about.

 

1. It is a SFF, air flow is very restricted.
2. Those temps are at 575/550 not stock.
3. Sampsa said that the the XT heatsink/fan was enough to get the XL at XT speeds and under good temps.

Hey, its not bad as people were suggesting.

 

well, in special cases it might be possible.
there are many possibilities of limitations, without exactly knowing where the limits are in d3 with their settings its hard to say, isnt it? someone should ask them to verity their results.

 

It is 15% and limited by the driver interface. Edit I thought you where talking about the XL overclock - my bad.

 

Where do you get "completely memory bandwidth limited"...? In some gpu architectures, just as in cpu architectures, you can increase the memory bandwidth without increasing the gpu clock and attain higher performance because the gpu runs more efficiently per clock with higher memory bandwidths (a la the P4.) I recall from overclocking my ancient nV25 that nVidia almost always clocked the memory higher than the core--whereas 3dfx often clocked them the same, etc., IIRC.

It's the basic difference between architectures designed for asynchronous and synchronous memory access. In a synchronous design you'd likely see lower performance by raising the mem clock above that of the gpu, but in asynchronous designs you'd likely see increases by raising the mem clock above that of the gpu. So, depending on the nature of the gpu architecture and what it is designed to do, you could well see a >10% improvement in performance by raising the mem clock by 10%, if the gpu core becomes more efficient per clock by doing so.

I think you might be mistaken is in assuming a "memory bandwidth limited" condition which doesn't take into account a possible rise in gpu efficiencies which can occur when raising the mem clock but without raising the core clock at all. You are assuming that the per clock efficiencies of the gpu do not increase when raising the memory bandwidth, it seems to me.

 

Where do you see the P4 scale super-linearly with the available memory bandwidth, all other things equal?

 

Could just reflect some bad settings in the drivers that control memory usage, etc..., No?

 

It might, but it's more likely to reflect bad benchmarking methodology.

 

I agree it's more likely to be the case of sloppy benchmarking methodology.

I thought I'd mention the driver configuration items, since DB mentioned what Orton said. It was something along the line of ''outside developers as well as their internal engineers haven't been able to do much with the previous batches of cards -- too inconsistent and flaky behavior''. We also saw on previous ATI hardware releases they achieved significant performance increases by tweaking/configuring how the memory subsystem is actually employed.