Recent Radeon X1K Memory Controller Improvements in OpenGL with AA

Geo

Mostly Harmless
Legend
Anybody seen the front page of www.hexus.net today? Wonder what it means. . .software or x1k hardware? If hardware opts for the X1k mem controller, how long to fold in significantly more apps? And might a few of those already been in place for X1K launch (like, say BF2 and FEAR)? Tho the NV 512mb card should give us some hint there.

HEXUS.beans :: ATI produce tool to increase Doom3 scores 'up to 35%' with AA enabled

Posted on Thursday, 13 October, 2005 by Ryszard

ATI Technologies have this morning made a tool available to HEXUS which supposedly improves scores in Doom3 when antialiasing is enabled. Improvements 'of up to 35%' are explicitly mentioned by sources within ATI. The tool seemingly changes the way the graphics card maps its own memory to better deal with handling AA sample data in Doom3's case.

The fix will shortly be rolled into CATALYST 5.11 according to ATI sources and a beta drop of that driver will be made available for testing in due course, before the final WHQL driver from Terry Makedon's CATALYST team is made available for public download in November.

HEXUS are in the process of testing the fix and we'll bring you some scores using the executable tool very shortly. It's unclear if the fix affects other games or whether it solely affects Doom3 via ATI's application detection scheme and their ability to reprogram the GPU's memory controller per application using CATALYST A.I.

More on the tool and the fix as we get it.

Edit: Now it's own page with graph: http://www.hexus.net/content/item.php?item=3668
 
Why not ask Charryszard Directly?

You've seen that a 512MB X800XL can beat the GTX in demanding circumstances *AND* you know that Doom3, even at UQ, doesn't *NEED* 512mb!
 
neliz said:
Why not ask Charryszard Directly?

You've seen that a 512MB X800XL can beat the GTX in demanding circumstances *AND* you know that Doom3, even at UQ, doesn't *NEED* 512mb!

Good point on that latter.

Well, now, as for His Hexusness, I imagine he'll see this and chip in as he sees fit and testing gives him much to say.
 
Haven't had much time to test it today, other stuff needs doing first, but yes, performance goes up.

31% @ 1600x1200, UltraQ, 4xAA, 8xAF (or whatever the UQ default AF level is).

More numbers later on, and analysis if it affects other games too.
 
Yeah, he's talking about the same thing that I have. It was pushed out with Doom3 as the poster child this morning, but actually paying some attention now, a tool name of "ATIAAMemMapSwitch" suggests it's not just D3-specific as I first wrote.

More coffee please.
 
Rys said:
Yeah, he's talking about the same thing that I have. It was pushed out with Doom3 as the poster child this morning, but actually paying some attention now, a tool name of "ATIAAMemMapSwitch" suggests it's not just D3-specific as I first wrote.

More coffee please.

Presumably, this only impacts the X1K series...but rather than presume, I'll ask. Does this only impact the X1K series?
 
Yessir, supported hardware for the little tool is X1K only.
 
Rys said:
Yessir, supported hardware for the little tool is X1K only.

Yeah, well, Dark Geo never 100% believes those things initially. Too many times in the past IHVs have tried to slide a software fillip off as "new gen only", only to be caught later by the bright boys and their reg hacks. I hate being that suspicious tho, and it would be way cool if this is legit for tweaking that controller.

Edit: What, if anything, does it tell us that it seems aimed generically at OGL? Does that say more about how they did it than what they did? Is it reasonable to assume that OGL apps have some generic similarity in their memory access patterns that doesn't apply to D3D?
 
geo said:
Yeah, well, Dark Geo never 100% believes those things initially. Too many times in the past IHVs have tried to slide a software fillip off as "new gen only", only to be caught later by the bright boys and their reg hacks. I hate being that suspicious tho, and it would be way cool if this is legit for tweaking that controller.

Edit: What, if anything, does it tell us that it seems aimed generically at OGL? Does that say more about how they did it than what they did? Is it reasonable to assume that OGL apps have some generic similarity in their memory access patterns that doesn't apply to D3D?

That was my suspicion too. Perhaps it is as you say. I'd like to know more about how this works first though. I suppose the first question might be whether or not they currently handle AA under opengl differently than they do under direct3d...

Nite_Hawk
 
Nite_Hawk said:
That was my suspicion too. Perhaps it is as you say. I'd like to know more about how this works first though. I suppose the first question might be whether or not they currently handle AA under opengl differently than they do under direct3d...

Nite_Hawk

The fabled OGL rewrite snuck in and marketed as hardware improvements? :p

Okay, I've worked off my nasty suspicious side for the moment and will go back to just hoping this is as yummy as the original reports suggest until/unless further evidence goes another direction.
 
Nite_Hawk said:
Did they happen to say what exactly it is doing?
As far as I was aware it tweaks some registry parameters that alters the configuration of the memory controller.
 
Joe DeFuria said:
Just to be clear...does this new memory mapping ability only specifically apply to the "ring bus" X1600 and X1800? (Opposed to the X1300 which lacks the ring bus?)
Unclear. AFAIK X1300 lacks the ring bus, but I think the controller logic itself might still have some commonality.
 
geo said:
The fabled OGL rewrite snuck in and marketed as hardware improvements? :p

Okay, I've worked off my nasty suspicious side for the moment and will go back to just hoping this is as yummy as the original reports suggest until/unless further evidence goes another direction.

Say, wasn't Humus hired to work on the OGL stuff at ATI? :devilish:

Dave Baumann said:
As far as I was aware it tweaks some registry parameters that alters the configuration of the memory controller.

That's a pretty impressive improvement for a memory controller tweak... Just to eliminate the obvious, color compression is currently active on ATI cards in OpenGL right?

Nite_Hawk
 
Just a couple of comments from Raja:

I saw a few posts...Note that running the tool on R4xx might bring the performance down a bit - since this tool is blindly forcing the driver down the path that's optimal memory-wise for 5xx. The driver deals with it correctly

Also, note that this is still not yet the optimal path for 5xx. This just better than what we shipped with.


Btw - This should improve scores for all opengl AA games..Not just Doom3.

- Raja
 
Quick Doom3 numbers from the XT only. Yay for coffee breaks.

bench-doom3.png
 
Back
Top