ATi 4xxx Series Review

[schneiderguy]

The clocks might be correct but the SP number isnt

 

[HurleyBird]

Originally posted by: bryanW1995sorry, but the 9800gtx+ is almost definitely going to be faster than 4850.

Without AA, yes. With 4xAA the 9800GTX+ should be pretty even, maybe a bit slower overall. Above 4xAA and the 9800GTX+ still gets raped.

 

[ChronoReverse]

Okay, I was skeptical about the 8xAA thing but from my own tests, it does appear to be applied.

Using my custom settings in WoC (because Very High isn't that smooth and I prefer smooth), the difference between 4xAA and 8xAA is about 2-3FPS... 49FPS versus 51FPS. Interestingly enough 2xAA is about 53FPS while NoAA is 55FPS.

Going from 4xAA to 8xAA costs less than 10%.


I've started to test the CFAA modes now too. Under the same settings, 12xAA (4xAA+ED) is about 40FPS. For 24xAA (8xAA+ED), it is 31FPS.

 

[HurleyBird]

Some of the card manufacturers thought the RV770 had 480SPs? AMD sure is paranoid these days -- justifiably so given NVIDIA's reputation.

 

[bryanW1995]

Originally posted by: HurleyBird

Originally posted by: bryanW1995sorry, but the 9800gtx+ is almost definitely going to be faster than 4850.

Without AA, yes. With 4xAA the 9800GTX+ should be pretty even, maybe a bit slower overall. Above 4xAA and the 9800GTX+ still gets raped.

yeah, I'm seeing more and more results that make me think that 4850 = 9800gtx+ and is ~ 10% > 9800gtx. however, it is by no means conclusive at this point as not many sites have up their results yet.

 

[ronnn]

Originally posted by: bryanW1995
What we really need is for somebody (steel six maybe?) with a 4850 to run benchies on a bunch of titles, have keys run the same benchies, then have the trade video cards and run the same benchies again.

Tons of benchmarks out there. More taken by a non-professional would just muddy the waters. I think by the time the 9800gtx+ hits the road - a winner will be clear to most of us. The maturing of drivers will likely play a role in this.

 

[Sylvanas]

Can't wait to get my hands on one of these and whack it on water.

 

[ronnn]

Originally posted by: ddarko
Examine Firing Squad's benchmark and it doesn't seem to square with other sites. Look at Bioshock: at 1920x1200, Firing Squad has the 4850 losing to the 8800GTX, let alone the 9800GTX and 9800GTX+. Look at Anand's results on the same game, same resolution: Anand has the 4850 absolutely crushing the 8800GTX (87.8 vs. 58.9); in Anand's benchmark, the 4850 even surpasses the GTX260 at that resolution.

The disparities between benchmark continues with Call of Duty 4. Anand, Firing Squad and Tech Report all tested this game. FS again has the 8800GTX beating out the 4850 at 2560x1600 resolution with 4xAA and 16xAF turned on. But TechReport tests at the exact same resolution and settings has the 4850 over the 8800GTX. In both cases, the difference was small, a couple of frame rates so maybe it's within the margin of error. Anand doesn't explicitly specify what settings it used at 2560x1600, although the chart says it used 4xAA at its 1920x1200 resolution so maybe we can assume 4xAA was also used at 2560x1600. In any case, Anand also has the 4850 getting higher results than the 8800GTX at 2560x1600 by a healthy margin.

So you're left with two sites - Anand and TechReport - converging on the relative performance of the 4850 versus other cards and Firing Squad presenting contrary benchmarks. I personally put more stock into Anand and TechReport than I do FS. Since I'm skeptical of Firing Squad's results in 4850 versus 8800GTX, I'm not going to rely on their benchmarks for 4850 versus 9800GTX and 9800GTX+ (FS also has the regular 9800GTX regularly outperforming the 4850). Anand and TechReport haven't tested the 98900GTX+ yet. I'm waiting to see them and other sites benchmark the 9800GTX+.

Always good to see a few sites, as results can vary. Also hard for the reviewer not get some unintentional bias mixed in. Anand used a fraps run through for bioshock, not sure what fs used.

 

[bryanW1995]

it actually wouldn't surprise me at all if amd had told all their partners that it was 480sp's. that would have contributed to the 1:38 am emails to anand and co and the whole 9800gtx + fiasco.

 

[paladiin]

So I picked this up today with the BB deal and have playing around with it all day. My biggest concern is the heat. With the Hotfixed Cats it shows mine idling at 79c-80c. However I noticed several things. First in the Anand review it shows the idle speed of the card at 160Mhz/500Mhz. Mine idles at like 500/650 if I recall correctly. Can't figure out how to set it to idle slower. Anyhow I also noticed that the "GPU Temperature" dial on Anand's and other reviews shows 55C at 1/3 of the dial (4th notch up). My graphical display shows 85C as the 3rd notch. Anyone else notice this difference?

Also, since I read a post saying that someone lowered their scores by 20-25C by using Arctic Silver I took my HSF off, cleaned the crappy stuff off, and put some AS on. My idle didn't change at all. However, I ran Bioshock for half an hour immediately followed by WoW for another half-hour with GPU around 65% constantly and my load temp is 84C. So I'm only getting 4C temp changes between idle and load. That can't be right, can it?

I really hate 80C idle temps, but I can live with 84C load for now I guess.

 

[ddarko]

Originally posted by: bryanW1995
it actually wouldn't surprise me at all if amd had told all their partners that it was 480sp's. that would have contributed to the 1:38 am emails to anand and co and the whole 9800gtx + fiasco.

It makes a good story but highly unlikely that ATI would have lied to its partners until the very end. Partners have to prepare manuals, boxes and advertising in advance. ATI deliberately tells them false information that then gets passed along to the public through the partners? That's a false advertising lawsuit right there, a stupid lawsuit, maybe, but a real lawsuit. Plus, pissed off partners who have to reprint everything that has wrong info in order to mitigate those lawsuits. All this trouble to hide info that will become public as soon as the product goes on sale? Very unlikely. Much more likely it's just a careless web page designer.

Originally posted by: ronnn
Always good to see a few sites, as results can vary. Also hard for the reviewer not get some unintentional bias mixed in. Anand used a fraps run through for bioshock, not sure what fs used.

Exactly. Everyone uses different systems and they test in different areas even when they test the same game so you can't expect that the actual numbers would match across sites. However, you would think the performance of the cards relative to each other would more or less line up across the different sites. That's not happening here, Firing Squad is the odd man out. It's possible FS is right and both Anand and Tech Report are wrong but it seems like more and more site match Anand and Tech Report's conclusions about the relative performance hierarchy of the cards than they do with FS.

 

[BFG10K]

Okay, I was skeptical about the 8xAA thing but from my own tests, it does appear to be applied.

Using my custom settings in WoC (because Very High isn't that smooth and I prefer smooth), the difference between 4xAA and 8xAA is about 2-3FPS... 49FPS versus 51FPS. Interestingly enough 2xAA is about 53FPS while NoAA is 55FPS.

Going from 4xAA to 8xAA costs less than 10%.

Very interesting, but have you taken screenshots to confirm there are IQ gains when going from 4xAA to 8xAA?

If so there must be some big architectural changes to allow such high AA performance despite having so little bandwidth to play with. It's possible the RV7xx can take eight multi-samples per clock cycle.

 

[ChronoReverse]

I have screenshots but you'll need to find me a place to host 2MB PNGs. If you want JPEG just let me know.

With that said, the difference between 4xAA and 8xAA isn't that big. I've always felt 4xAA was a sweet-spot especially with progressively higher resolutions.

Here's a peek at where the AA does make difference and how small that difference is.
http://img145.imageshack.us/im...756/4xaapreviewln0.png
http://img145.imageshack.us/im...299/8xaapreviewmb7.png
http://img80.imageshack.us/img...50/24xaapreviewbh2.png

 

[BFG10K]

High quality JPEGs will be fine. Try to get something with angled edges at a long distance and I'll use zoom to verify there's a change.

Also just to confirm that you?re using box filtering for both 4x and 8x? Because if you?re not then you?re testing CFAA which is something else entirely.

Very exciting times indeed if true; this could be the architecture that establishes 8xMSAA as a standard like other architectures established 4xAA so many years ago. :thumbsup:

 

[BFG10K]

Sorry, I missed your edit. You'll need to redo the images as high-quality JPEGs as Image Shack has scaled them back because they're too big.

 

[ChronoReverse]

Those two are just tiny clips I made that show one area where the difference is apparent. The two screenshots are taken from exactly the same camera angle so that I could compare.

I'm going to have to try to find another place to take a screenshot that shows edges better.


And yes, I'm aware of the difference between the different filters =). I have the results for the ED mode too.

 

[qliveur]

@BFG:
Your statement about 8xMSAA being the new standard makes sense in that ATI would optimize it's new architecture for 8xMSAA since it supports DX10.1 and one of the requirements of DX10.1 is mandatory 4xAA (I think).

Am I making any sense here?

 

[BFG10K]

ChronoReverse, it's cool, I can see a definite difference with 400% zoom; 8x is way smother than 4x, especially the bottom edge. Just confirming you used box filter for both images?

qliveur: I don't think the DX spec has anything to do with it, I think it's ATi trying get back into the IQ game after being behind for so long.

 

[SniperWulf]

Originally posted by: GTaudiophile
Looking at this 4850 vs. 9800 GTX+ review at FiringSquad, the AMD card doesn't fair too well in Crysis...

http://www.firingsquad.com/har...ce_9800_gtx+/page6.asp

Hell, neither of the cards tested fair well in crysis (the most poorly coded game of 07)

 

[ChronoReverse]

Yes they're both set to box filter.

And just for kicks, I've added the 24xAA screenshot (REALLY minimal now).
http://img80.imageshack.us/img...50/24xaapreviewbh2.png

At 800% I can tell the difference from 8xAA, but it's really tough otherwise.


Did some more tests at 4xAA with the various filters.
Box: 51
Narrow: 45
Wide: 45
Edge Detect: 40

Just to hammer this in, 8xAA is 49FPS.

It is cheaper to do 8xMSAA than to use CFAA

 

[HurleyBird]

Wow, the difference with 8XAA is huge. Hard to believe the performance difference is so small.

 

[ChronoReverse]

I used a tool someone told me to try at Beyond3d to confirm 8xAA is indeed active.

8xAA
http://img78.imageshack.us/img78/6650/8xaavr4.jpg
4xAA
http://img131.imageshack.us/img131/1218/4xaasz7.jpg
2xAA
http://img78.imageshack.us/img78/8638/2xaagv8.jpg
NoAA
http://img113.imageshack.us/img113/9651/noaalx5.jpg

 

[lopri]

Thanks for the screenshots, ChronoReverse.

Has anyone seen this? (First video from the link)

http://www.amd.com/us-en/Corpo...06,00.html?redir=cin01

Realtime??

Edit: Bigger (and blockier) version

http://www.amd.com/us-en/asset...Media/AMD_Ruby_S04.swf


P.S. Oh and I found this useful pic.

http://i114.photobucket.com/al...3/Wirmish/GPU-Dies.png

 

[schneiderguy]

For comparison, heres 0x AA vs. 4x AA vs. 8x AA on a 3870 in Crysis. Sorry no screenshots, they turn out completely black when I use the printscrn button

Crysis high settings, first level, looking at some rocks and trees:

0x AA: 36.3fps
4x AA: 24.2fps (33% performance decrease)
8x AA: 21.6fps (10% performance decrease)

*shrug*

Hardly any sites test with it on except for computerbase which show the 3870 beating the 8800gt in most tests with 8x AA enabled (and in a couple of games it even is faster than the 8800gtx/ultra)

 

[Sylvanas]

Originally posted by: lopri
Thanks for the screenshots, ChronoReverse.

Has anyone seen this? (First video from the link)

http://www.amd.com/us-en/Corpo...06,00.html?redir=cin01

Realtime??

P.S. Oh and I found this useful pic.

http://i114.photobucket.com/al...3/Wirmish/GPU-Dies.png

That video is pretty neat and was rendered in Real time. Here is an article explaining AMD's 'Cinema 2.0'. That die comparison is a good resource for later too thanks Lopri, bookmarked.

I am pretty sure I mentioned a Rage3d thread a while back where it was discovered the 3870 took a small performance hit with 8xAA- and here we are now, seems the extra changes in the RV770 improves on that again.