radeon 3870 lose at 4AA/16AF against 8800GT but win at 8AA/6AF?

As pointed out already, those are the %'s with HD 3870 being the base (100%). In that review, if you scroll over the graph, it changes from fps to %.

 

Thats a very interesting review

Reading other reviews I found it odd where fairly often they would quote something about how RV670 got slaughtered at AA when in fact their results showed the 3870 besting the GT at the highest resolution with most AA/AF enabled.

So some stuff seems to have changed indeed in RV670.
Dave previously mentioned some stuff improving latency sensitivity.

An earlier thread (edit: this one) pointed out that cache hit improvements etc. tend to make the AA/AF workload proportionally less with higher resolution.
Maybe the HD bit of marketing is actually legit & R600 architecture is better suited to higher resolutions than the NV architecture.

A much older R600 thread included suggestion that the 64bit memory channels of R600 would probably often not be fully utilised resulting in inefficiency.
A much higher clock rate 256bit bus made of 32bit channels (eg RV670) thus may actually have much higher usable bandwidth.
Presumably absolute (but not necessarily per-clock) latency is better with both higher core clock & memory bus clock too?

The few OC benchmarks at 850+ I've seen seem to look much more like direct competition to GTX.
(Are ATI guys allowed to disclose the design target clocks for R600 since its now EoL?
Or while we're on it, the rationale for the 512bit bus?)

 

Funny, I'd forgotten that I suspect RV670 has 32-bit memory channels - which should have an effect on latency. Dave's also mentioned latency related changes but we don't know what they are.

Jawed

 

It would be useful to conduct a handful of tests with GDDR4-based R600 board and one with RV670 (the HD3870), with similarly clocked memory arrays and GPUs. That would shed some light on the matter.

 

Sure it is not playable, but think about a second card: 3870 CF for ~$450 would beat 8800Ultra SLI for ~$1200, while reaching 30FPS+...

The problem is definitely on NVs site, FPS divide by four from 4xAA to 8xAA, what should not happen...

I see 3870 drop from 4x AA to 8xAA in FPS: 31 -> 21. NV should drop equal, but like I said FPS divided by four...

 

Do both vendors use the same method of AA at either 4x or 8x?

 

The fast answer is yes, they both do gamma corrected Box AA by default at 4x and 8x.

The not so fast answer is that both vendors also have differing forms of 4x and 8x and everything in between and beyond.

Nvidia also has dedicated hardware resolve for AA. You'll notice lots of hoopla when something doesn't use the hardware and instead uses "software" or "shader based" AA (IE - COJ DX10) to avoid rendering errors when AA is enabled.

ATI uses shader based AA with hardware dedicated to a "quick path" from RBE's to shaders for this purpose.

Maybe this is just a case where Nvidia hardware is optimized for 4xAA and as you get to higher levels of AA it equalizes out or is less efficient than ATI's shader based AA. Who knows...

Regards,
SB

 

Who the hell buys a 8800U to run stuff at 1280x1024 ? Numbers for top end cards on anything less than 1600x1200 are pointless IMO. Ive not used less than 1600x1200 in 8 years.

 

There are a lot of users with nice 19" LCDs with 1280x1024 native resolutions. I have two such displays in dual display.

And shock and horror, in a game like Crysis I would keep all the settings on high and go down to even 1024x768 (or even worse, 1024x576) to get stable framerates where 90-95% of frames are above 30fps. I used to game on a 1600x1200 21" CRT, and as nice as it was, I think resolution is overrated when it comes to compromising features (notably AA and AF) or framerate.

 

So can we pinpoint why the 8800GT loses more performance against the 3870 when 8xAA is enabled?

Bandwidth? framebuffer? drivers?

However, i wouldn't be surprised if 16xCSAA results are much better than the pure 8xMSAA mode. Infact, 16xCSAA is known to not only provide better IQ, but the performance impact is roughly equivalent to 4xMSAA.

Reviewers should use CSAA as well in these cases IMO. Why? because these modes dont hit hard on bandwidth/framebuffer (the very concept behind CSAA) like the traditional modes. (Look at the performance hit of 8xMSAA on 8800GT!)

 

CSAA doesn't provide consistent quality across all games (compatibility) and polygon edges (polygon intersections).

Better compression and ring-bus?

 

I'm willing to bet that it's something as simple as the fact that G92's hardware accelerated/assisted AA is optimized for 4xAA with little thought given to 8xAA.

And when moving to 8xAA it has to spend an inordinate amount of time rendering it or maybe it's falling back to shader based AA at that point.

Since Rv670 doesn't have dedicated AA hardware (other than the fastpath to the shaders) it has more of a linear-ish drop off than G92, and with the abundant ALU power, I don't imagine 8xAA is going to be inordinately more taxing tha 4xAA.

Regards,
SB

 

I think its bandwidth, here is the test continue with 3850, 3870 at 1280x1024 21.9% faster than hd3850, with 8xAA 16xAF enabled 35.4% faster, 1600x1200 hd3870 26.2% faster, with 8xAA 16xAF 56.2%.

3870 8xAA 16xAA performance looks really great against competation, the problem there is only very few game can the card run with playbale framerate at 8xAA 16xAF.
(the thread name is wrong its say 6af )

 

I've always assumed it was the infamous "memory management bug" on G8x/G9x hardware, rearing its ugly head again. Given computerbase.de's numbers showing the 3870 beating even the 8800 Ultra w/8xAA @ 2560x1600 across a wide variety of games. Since the Ultra has more bandwidth than any 3870, I'm going to go ahead and say it's not a bandwidth limitation

 

How can you say that's not a bandwidth limitation, when both solutions use different compression techniques?

 

Because one has way more bandwidth than the other yet yields lower performance.

If we believe it is a bandwidth limitation on the NV side, then why does the Ultra with ~50% more bandwidth than an HD 3870 lose out to it at such a high resolution with 8x AA across so many titles (12% aggregate)? If it is bandwidth, then a G8x/G9x product would need something like 70% more bandwidth than an HD 3870 to recapture the performance crown with these settings, and I just don't buy that. Also, if it is bandwidth, why doesn't the HD 2900 XT with its massive 100GB/s+ bandwidth out-perform G8x/G9x by even more?

It's not bandwidth.

 

It's not bandwidth in the AMD/ATI camp.

It is bandwidth in the Nvidia camp.