AMD: R7xx Speculation

Believe or not World of WarCraft was the poster child for me when investigating performance and IQ with adaptive and TA features from ATI and nVidia.

The Alpha test art work is so diverse with some areas with so many alphas, well, it's daunting.

FarCry was another title with a lot of alphas to investigate specific scenes but the alphas were very similar and a bit tougher to gadge quality aspects.

Serious Sam 2 in certain areas was a nice Alpha-test, test candidate.

Actually there are quite a few titles.........always seem to investigate the real heavier scenes.........cupcake examples are kinda worthless.

Never could find real viable data on Alphas -- an occasional bench and a quickie investigation on quality and that was about it -- too cookie cutter most reviews are to me.

 

http://cid-1b1908eb94064033.skydrive.live.com/self.aspx/Public

TSAA results included. Taken at 4xAA/8xAF, 1680x1050



The reason I think something's wrong is because between the various modes, I can't see any difference nor are there any FPS differences. Can anyone tell what mode is being used? Something is definitely being done, but between TSSAA and TMSAA there doesn't seem to be any change.



I don't own any of those games. Is there a demo with such scenes available?

 

I don't have CCC installed so I can't verify what settings it's using. I can see that adaptive AA is working, though, as seen in the tree in the HL2:Ep2 shots.

 

You can see some affect on the trees in WiC as well. Though I can't tell (much of a / any) difference between the TMSAA 2 / TMSAA P / TMSAA Q shots.

 

That's what I meant.

Obviously the Adaptive AA is working but I can't tell any difference between the three modes. There's no performance difference between them either.

Notice that the penalty for turning on Adaptive AA is very low as well.

 

How's the penalty for Adaptive AA with 8xMSAA ?

 

4xAA 108FPS
8xAA 101FPS
8xTSSAA Q 91FPS

Screenshots in the SkyDrive again.

http://cid-1b1908eb94064033.skydrive.live.com/browse.aspx/Public

 

Ouch? Another price drop on the 260?

 

Looks like the Chinese have the 4870 reviews up. Impressive:

http://translate.google.com/transla...167d9273.html&hl=zh-CN&ie=UTF8&sl=zh-CN&tl=en

 

Looks to be even or better with the 260 aside from lost planet.

 

It's quite apparatent in NfS Carbon, Quick Race, Francis Tunnel with aniso forced via driver or in HL2 Episode 2 in the beginning (Into the White Forest?), especially with installed Cinematic Mod 8 which addes very lovely and highly detailed textures to the game. But also in the railways in "route canal" as i believe it's called.

For NfS Carbon it's this spot/track:
http://www.pcgameshardware.de/&menu=browser&image_id=740631&article_id=621293&page=3&show=original
(an older image without HD4850 still)

But then, Nvidia is - apart from the mentioned problems with track-markings (not sure which kind of texture-function to attribute that to) in most racing games - not perfect either:
http://www.pcgameshardware.de/aid,6...ltering_quality_despite_of_HQ_driver_setting/

I agree (while also agreeing on the fact, that comparing anyhting else than feature-test results from Futuremarks' products make less sense with every iteration).

But then, if their goal was to measure gaming performance, they (FM) direly need to integrate multi-core support into their game tests as well as physics (note - not with an "x") with multiple APIs (PhysX & Havok & ODE) because all the games i know and play do not have separate levels putting strain only on either GPU, CPU or "P"PU.

The 'Problem' (actually, it's not a problem, but a feature), is that HL2 EP2 (blablabla) in my tests with Cinematic Mod 8 actually did some transparency-AA on the fences all by itself as soon as you select FSAA in-Game. But it seems, trees an stuff further back in the distance does not profit from this. So on those objects you still see improvements from driver-applied transparency AA - but you also see, that the current (well, Cat 8.5 was current, when I did this) drivers do not make a difference in HL2 whichever adaptive mode you select.

 

merged with above, pls delete

 

If those numbers are true I am definitely getting the Radeon HD 4870!

If only it came with an EFI compatible ROM

This thing is a monster at 1920 x 1200.

 

Gebus. If those numbers are anything close to reality, the 4870 is generally faster than the GTX 260 for less.

Probably why there's rumors of ATI increasing the price to 349 USD. Crossing my fingers that's not true.

Either way, looks like it's going to be a long summer for Nvidia.

Although, oddly enough, right now I have to give Nvidia more credit for a lower idle power than ATI. I'm hoping it's a driver bug that isn't reducing the core frequency and memory frequency as low as it should be. Notice that Rv670 drops to 300 mhz on the core when idle versus Rv770 only dropping to 500 mhz on the core when idle.

Still. Pretty damn impressive. In most cases 4870 is roughly twice as fast as 3870.

Regards,
SB

 

No, final MSRP is $299.

But it seems availability could increase prices to GTX 260 level:
http://forums.vr-zone.com/showthread.php?t=292880

 

Performance is great, but power consumption is troubling.

It uses 33 Watts more in Idle compared to the 4850, which is strange because one would assume that it decreases frequencies and voltages to the same level and it's the exact same chip.

It uses 58 Watts more under load compared to the 4850, which is at the same level as a GTX280.

I doubt that a straight 4870X2 card would be possible with these power characteristics. They will probably need to decrease frequencies.

 

I wouldn't draw any conclusion based on one measurement. Different games/benchmarks and settings (ALU:TEX, different limitations...) leads in results, which can differ in tens of W.

 

Indeed, Furmark should be the application, which should be able to utilize all 800SPs.

 

FWIW, some claimed 4870 performance pictures from an AIB maker:

http://www.force3d.com/press/23_6_08_hd4870.html

X3321, P7534.

 

NFS has a history of using negative LOD bias, so it could just be that NVidia clamps/ignores it while ATI is respecting the dev's setting.

Devs using negative LOD bias should be shot, though.