The NEXT LAST R600 Rumours & Speculation Thread

Notable items:

• Re-Z
• Z Range optimisation - one of the patent applications described this, which is a way to improve Z-testing in the hierarchical-Z system, I believe
• Memory read/write cache for improved stream output performance - not surprising, good to see the theory confirmed
• Bicubic texture filtering - someone suggested this when I dug up a patent about texture filtering, seemed unbelievable
• PCI Express x16 bus (seems to imply it's not PCI Express 2.0)

http://forum.beyond3d.com/showpost.php?p=972191&postcount=2363

Andy suggested bicubic filtering a while back. This is the relevant patent:

http://forum.beyond3d.com/showpost.php?p=972154&postcount=2362

Jawed

 

From what I know 8.37.4 are the latest and the review drivers.

 

Slightly?

1600x1200 4xAA/16xAF
Prey: GTX 40% faster
Fear: GTX 26% faster
Quake 4: GTX 13% faster
X3: GTX 10% faster

Sorry, there is no way to call this "slightly".

 

By this W1zzard dude who posted the review.

http://www.xtremesystems.org/forums/showpost.php?p=2182156&postcount=14

By the way, did anyone happen to see what system he used to bench the cards?

 

So, no hope for Lanczos, then.

 

Yep:

 

I ment that, if DDR3 is enough for the GTX, why would it hurt the XT? the Hello 8800 GTX was ment as "here's another product that's running perfectly fine on DDR3"

 

How much a GTS scores at these settings? (and remember that you are talking only about the worst score at every resolution, a fair comparison is an average - and at very high resolution + AA, there's a possibility that XT it's also limited by frame buffer size. )

 

I'm wondering what the performance will be like. Will it be usable?

Also, is it on top of trilinear/AF or is it instead of? That's my stupid question for today.

Jawed

 

Thanks.

 

It's a planar image resampling (filtering), just like bilinear & etc., it's nothing to do with the higher order of texture interpolation, IMO.
It's costly, though -- quite a lot of MULs and trig op's per tap.

 

The 8800 GTS isn't listed in the various tests. Very crappy imo, because the main competitor of the 2900 XT is the GTS, or perhaps the X1950XTX? We'll see. If these numbers are real, I dunno what ATI did the last months. The spec of the 2900 XT seems fine, but even the predecessor can compete with this new, high-tech card. Creepy...

 

The GTS is right around the 1950xtx other in games that have heavy shader usage where it pulls ahead by quite a bit. (oblivion is a good example of this)

 

So, whenever a bilinear operation is performed, bicubic replaces the bilinear operation? So this works only for bilinear and trilinear filtering?

Jawed

 

That's beginning to worry me a bit, too. You'd have thought the 512-bit bus would allow it to narrow the gap between it and the 8800GTX as the resolution/AA/AF goes up; but if anything it seems to fall even further behind. Could it simply be running out of memory when working at very high resolutions? Is 512MB not enough?

 

I wonder, if this bicubic resampling is a "full" implementation, does it [the hardware] support programmable coefficients -- say for more blurry or more sharpen output?
It could be somewhat "locked" to a MIP level (AF?) and with gradually iteration to sharpen the output texel as the MIP level increases to aid the AF resampling. D'oh!

 

Yes, what I see is that R600 is a card that on paper has way more power of X1950XTX. 25-30% more shader power, texturing power unknown but likely to be at least 10-15% more (if they kept the same R580 units, but they say they are improved). Beefed up ROPS and the bandwidth to feed them.
So, if it performs in average on par with X1950 XTX, it can be only

1) ATI made some terrible mistake in designing R600, and there are bottlenecks and chip problems impairing the performance, (i.e. trying to put too many features on it but missing some resouce in one or more fundamental points) or
2) making a performance driver for R600 is really hard, with very big difficulties due to co-issue erformance penalties, practically needing the driver to be optimized heavily for each game or
3) both

And what makes me wonder were the "Preliminary" watermarks all around...

 

hmm that sounds mighty familiar

 

OK, THIS is undoubtely a joke

http://www.techpowerup.com/reviews/ATI/Radeon_HD_2900_XT/1

 

Or, both the TUs and RBEs are completely new, as well as the ring-bus being heavily revised - and they really haven't got on top of performance for these new parts.

I'm certainly not excusing them - they've had bloody ages to sort this stuff out.

Jawed