R520 info thread #(9!)^9

Bleh and double Bleh. Totally unimpressive. The only REALLY nice thing is the PS Dynamic Branching; WOW, just WOW! And in case that wasn't obvious, it's the reason Farcry is so damn fast on the R520, I would suspect. Sounds like NVIDIA's insistance regarding dynamic branching has finally payed off... although perhaps not in the best way imaginable for them

As for the higher AA/AF performance, I'd rather compare the GT to the XL here, and this highlights a quite important fact imo: without its 1.5Ghz memory, the R520's AA/AF efficiency improvements are in the middle single digits (7-8% average, I'd say). Not bad, but nothing NVIDIA can't beat with a 500+/1500 G70.

Still, congratulations to ATI for making Pixel Shader Dynamic Branching a kickass feature, and one that can easily be used without thinking too much about it ("hmm, is my average batch size above or below 763 pixels?") - now if only that was true of the G7x too, developers all around the world would be crying of joy. Oh well, at least it's a good indication that it's going to be an extremely useful feature in the DX10 timeframe.


Uttar

 

Overall impressions:

Love the no-holds-barred AF, gobs of fast memory and wicked fast dynamic branching performance (does F.E.A.R use this for soft shadows?). Everything else is really bleh though considering it will be out in November.

I'll get excited about the ass raping the GTX got in some newer titles at max settings once I see GTX results with 512MB of similarly fast memory - until then we won't know how spectacularly efficient the memory bus is. Also, the shader efficiency isn't all that hot since it gets killed by a stock GTX in SM2.1 and the 3dmark05 PS tests. Not to mention OpenGL and Quake4 still belong to the green machine. Avivo got trounced by Anand and H.264 support isn't coming anytime soon so that's a disappointment to some (not me since I don't do video on the PC).

Color me unimpressed for now - until the R580 or HDR+AA info rolls around.

 

Truly, you are the Frying Pan of Knowledge. (It was a joke anyway.)

As an aside, no reviews use 3DM05? The times, they are a-changin'.

 

Well now, that ati flash site for K1K is nice.

http://ati.com/designpartners/media/edudemos/RadeonX1k.html

^ Assassin and Toyshop demos in 720p. Toyshop was mentioned by a review that I cannot remember now, one in the x1800 reviews thread.

 

So AA + HDR does not impress you ?

Heck, I find the X1800's architecture quite innovative and feature rich. It bests the competition, while it provides the same performance.

Optimizing the memory controller for titles sounds good, too (I already can hear the exe-renamers and and "cheat" shouters ... )

 

We haven't seen AA+HDR yet - when we do I'm sure we'll be impressed. And like Uttar pointed out above, the memory controller doesnt seem all that efficient since the 7800GT is faster than the X1800XL in most games across the reviews I've read so far.

The only thing compelling so far is AF quality. The performance just isn't there to justify the late arrival.

 

Macci was using WinClock to bump clocks, so I've been trying to extort that out of someone, but it didn't work in time. I don't know of anything else that can adjust clocks on the hardware right now.

 

Are you saying that this card doesn't have a simple way to overclock right now? With all the rumours of high overclockability, that's like dangling a carrot on a stick ><

 

This card is months late. Would have been much more impressive this summer. Right now, it looks like ATI dropped the ball. It'll perform better across the board with better drivers and a 512MB card IMO. But this isn't the leap the delay should have afforded. PEACE.

 

NVidia's insistence on dynamic branching? You're having a laugh. NVidia's architectures are shit at DB for performance. It's purely a programming simplification, nothing to do with performance.

And, anyway, there is no dynamic branching in FC. Crytek said so at GDC.

Jawed

 

Dont forget thet that GTX has a fill rate advatagne thanks to 8 extra pipes (which means more PS power) that really help it over come the faster clocks of the 1800xt

 

EDIT: Oops, posted in wrong thread.

 

In their effort to appease nVidia back into being a partner again, they went too far and appeased their way into irrelevance. 3dMark03 was one of the better and most useful synthetic benchmarks ever; 3dmark05 is... something else.

 

They're shit at it if and only if the batches aren't coherent. "Insistance" might be exagerated, I agree - but considering their crap performance with it as you said yourself, they really don't discourage its usage so much in developer performance presentations - there always is the "coherent" tag up with it, of course, but still.

I checked, and indeed I stand corrected: it's only static branching. That makes me curious about R520's static branching performance compared to R420/G70 now, actually


Uttar

 

Bloody hell, has anyone seen the Anandtech review, they've really outdone themselves this time. Talk about selective benchmarking! Surely, this has got to hurt the site in the long run...

 

They (anand's) at least have a suggesting title so you don't have to read it at all.

 

Does it matter? Static branching can just be resolved by the compiler...

 

Delays don't allow leaps. Careful and smart planning does.

 

http://www.techreport.com/reviews/2005q4/radeon-x1000/index.x?pg=12

 

Yes, by a whole 3% (10320vs10000). Consider that the 7800GT and X1800XL have the same theoretical fillrate.