24xAA

[willhub]

From the experience I have had in games, the AA problem is still here and 2xAA - 4xAA gives too large of a performance hit, 24xAA is hardly playable on anything, totally pointless.

 

[Dutch Guy]

What are you on about?, have you checked some of the latest reviews. Fair enough performance is down in one or two games but it's mostly up there with the GTS and beating the GTX/Ultra even with AA enabled somtimes.

Performance is good compared to the 8800GTS/GTX but even you must acknowledge that the performance hit of the HD2900XT is bigger than Nvidia's 8800 cards when using FSAA.

I still feel the card is too slow if you look at the hardware specs, it should have a lot less impact from FSAA, even if it is done in shaders.

 

[Tom|Nbk]

Performance is good compared to the 8800GTS/GTX but even you must acknowledge that the performance hit of the HD2900XT is bigger than Nvidia's 8800 cards when using FSAA.

I still feel the card is too slow if you look at the hardware specs, it should have a lot less impact from FSAA, even if it is done in shaders.

True the hardware specs of the 2900 pretty much trounce the G80. Only problem is


A. G80 Has a shader clock 2900 does not.

B. G80 Does the AA resolve totally away from the shader units 2900 does not.

Im curious as to whether its possible in drivers to move the AA resolve to the ROP's unit in the card. I believe it has 16 Rop units similar to the X1950XT-X

 

[RavenUK]

Bah, i play with no AA at all and the performance is sweet, having a 20 inch monitor at 1680x1050 means the dot pitch is real small so jaggies don't bother me much at all just don't sit three inches from the screen, its like LCD and plasma tvs, the further you sit form it the better the picture gets.

 

[ergonomics]

Bah, i play with no AA at all and the performance is sweet, having a 20 inch monitor at 1680x1050 means the dot pitch is real small so jaggies don't bother me much at all just don't sit three inches from the screen, its like LCD and plasma tvs, the further you sit form it the better the picture gets.

then that defeats the point of a parge screen then doesnt it

 

[fornowagain]

True the hardware specs of the 2900 pretty much trounce the G80. Only problem is


A. G80 Has a shader clock 2900 does not.

B. G80 Does the AA resolve totally away from the shader units 2900 does not.

Im curious as to whether its possible in drivers to move the AA resolve to the ROP's unit in the card. I believe it has 16 Rop units similar to the X1950XT-X

They're two very different designs and not at all easy to compare directly. The R600 does have more functions, tessellation for example, but the basic SP are both very complex. Arguably one of the R600's biggest 'problems' is VLIW coupled with a superscaler architecture which can be very hard to program and horribly inefficient. But its lack of hardware resolve is certainly a problem if you like AA.

I'm not entirely sure what you mean by B. It's the resolve people talk about when they say the hardware is broken, obviously the R600 already uses render back ends (ROPs) for AA.

 

[Tom|Nbk]

They're two very different designs and not at all easy to compare directly. The R600 does have more functions, tessellation for example, but the basic SP are both very complex. Arguably one of the R600's biggest 'problems' is VLIW coupled with a superscaler architecture which can be very hard to program and horribly inefficient. But its lack of hardware resolve is certainly a problem if you like AA.

I'm not entirely sure what you mean by B. It's the resolve people talk about when they say the hardware is broken, obviously the R600 already uses render back ends (ROPs) for AA.

Nope the 2900 does not do any AA resolve in the render back ends (ROPS) afaik its ALL done in the shader units and the rops are used for z-buffer and a few other things.

 

[Gerard]

Nope the 2900 does not do any AA resolve in the render back ends (ROPS) afaik its ALL done in the shader units and the rops are used for z-buffer and a few other things.

Still dunno why they went that route knowing it would cripple fsaa performance.

 

[Tom|Nbk]

Still dunno why they went that route knowing it would cripple fsaa performance.

I think they wanted to take a risk and see if they could get away with the new arc, obviously they couldn't really

 

[fornowagain]

Nope the 2900 does not do any AA resolve in the render back ends (ROPS) afaik its ALL done in the shader units and the rops are used for z-buffer and a few other things.

Yes I wondered if thats what you meant. In fact it does the 'AA' which means samples in its ROPs, its the filters that are applied in the shaders. The resolve is where AA samples are combined and processed to a single point. A lack of hardware resolve means it has to combine/process the ROP samples using software (which is silly distinction as its still ON hardware, just not dedicated). The difference is the G80 has dedicated HW to do this function. Its a chicken and egg thing, you can't take AA samples until they have been created in the SPs. The ROPs grab the SP outputs, take samples, then has to loop the results back to the SPs for processing. Very hardware costly and its not something I believe they would have done out of choice, hence why some say its broken, its a performance killer.

Remember the fuss over Call of Juarez, its because they made the G80 use software resolve and not its normal HW resolve.



http://www.beyond3d.com/content/reviews/16/10
http://www.beyond3d.com/content/reviews/1/10

The ROP stage is likely going to be one of the most controversial parts of the R600 design for some time to come, given how AMD are making use of it in the shipping configurations. Basic abilities first, the ROP now supports 8x multisample AA using programmable sample grids, testing 4 positions per cycle with samples laid out over a 4-bit grid (16 discrete positions per pixel, X and Y, 256 total). It's able to test any supported surface format, including float surfaces. That means the hardware's basic multisample abilities match (and exceed by virtue of programability, you could argue) that of NVIDIA G80.

Even for the basic box filter resolves, where the hardware weights samples based on their proximity to the pixel centroid, R600 will perform the calculations to resolve the samples on the shader core, that is unless compression for the tile is at maximum, so you know the resolve would just return the same colour anyway, so there's no math involved to filter. Currently that points to the hardware resolve either being broken, at least under some conditions (when compression is less than maximum), or it being easier to maintain in code if you're doing other filters anyway, so you only have one path for that in the driver. We lean towards the former, rather than the latter, since the performance deficit for shader core resolve seems to be significant, even for the basic box filter cases.

 

[Tom|Nbk]

So what you thinking Forn, how versitile is the AA resolve on the R600 how much of it can and can't be tweaked since you obviously know a little more than me

 

[fornowagain]

So what you thinking Forn, how versitile is the AA resolve on the R600 how much of it can and can't be tweaked since you obviously know a little more than me

Its totally flexible and programmable, which is great and allows for good custom filters and better IQ for high AA on a per game basis. Trouble is, this time round, using the shaders will always hammer performance. I guess it comes down to how much can you optimise a filter anyway? The G80 HW filters are fixed, but the IQ's pretty good and doesn't take the hit. The other thing is how objects have AA applied, anything that overrides the game drivers runs the risk of applying aliasing to objects that don't need it or not applying it to objects that do. For both cards performance wise the basic MSAA (box) filter gets used most in game and should take less of a hit, but then you're stuck with the max 8x on both cards. Personally I think programmed filters is the way they'll do it in future, just on dedicated silicon. So its a step in the right direction and a step back for performance.

 

[Gerard]

Beyond 3d lol

That site should have some kinda slogan probably along the lines of "no matter how good your card is we think we can do it better" armchair engineers unite!!!

 

[fornowagain]

Lol shouldn't it just, its always easy to criticise after the fact. Still they do post the most information on the actual design mechanics.

 

[Tom|Nbk]

Forn you free to come on MSN some time I need to pick your brains about somthing . Sorry for going OT

 

[HangTime]

Surely the proper solution to jaggies is ultra-high resolutions?

Yep, this is what I've always said since AA first reared it's head on the gf2 range. AA is effectively a substitute for insufficient display resolution. If we had monitors (of current size) running 3200x2400 or similar we wouldn't need anywhere near the same level of AA, however such screens don't exist. There's also a case for 'intelligent' AA being more efficient than an increase in resolution, if getting rig of jaggies in the primary goal.

 

[Gerard]

Last i read increasing res does minimise the jagies by making them smaller but theres also more of them to contend with because of the increased res.

 

[Tom|Nbk]

So is it possible drivers can help the TMU/ROP short coming with the R600?.

 

[Dutch Guy]

So is it possible drivers can help the TMU/ROP short coming with the R600?.

Drivers cannot solve a hardware issue, or can drivers make a HD2900XT suddenly have twice the ROP's?

AMD made a decision to go with what is currently the HD2900XT, either because they had problems and were forced to or they took a gamble that (so far) hasn't paid off.

 

[mrk]

No the highest 'real' AA is 8xQ which is 8xMSAA. The next one is 16x which is 4xMSAA+16 coverage samples. After that its 16xQ which is 8xMSAA(8xQ)+16 coverage samples. They're composites, not real AA. You'll also notice that moving from 4x to 16x doesn't take as big a hit as you'd think, thats because its the same base sample.

I guess they call it 16x cos 4xMSAA+16CS is a mouthful.

Here's my results with 16xAA (standard) and 16xAF with everything on Max at 1680x1050

How can people tell the difference between 4xAA and 16xAA, I mean, 4x normal FSAA looks amazing as it is...