Ati Crossfire capable of 14X FSAA

[trinibwoy]

Pete & Xmas - appreciate the knowledge

 

[repi]

They should focus on making MSAA work with HDR rendertargets instead of just increasing the number of samples on LDR surfaces, which IMO not many next-gen games will use

 

[martrox]

It would be a pity if this "out of the park" IQ requires more than one video card because I think you'll find that the most fervent ATI supporters think using dual cards is absolutely stupid and pointless

On further reflection I gotta point out that the reason most "fervent ATi supporters think using dual cards is absolutely stupid and pointless" is that there is no need/huge advantage to it yet.

If you could show them an advantage to having dual cards I think they might reconsider their opinion, and 24xAA is definately looking to be an advantage.

Just a thought.

Absolutely, DW. Most of the problems I've had with SLI has to do with the fact that it's pointless to consider unless you go 6800GT/U. When a single 6800GT will give you pretty much everything for a lot less than 2 6600GTs, it's still a nobrainer. And all you get from SLI is a speed bump - IF it supports the game you want to play. The real question here is can ATI go nVidia much better on game compatibility. As a self admitted graphics whore, it's enough for me to consider.......show me better IQ and I'll go there......

 

[KimB]

Heh, "focus?" None of these things are remotely challenging to implement. It's about design decisions.

But mostly I agree, though, right now it's more important to get MSAA working at all with FP rendertargets than it is to get today's already good MSAA to look better.

 

[Dave Baumann]

Why? The only performance problem I am able to see at the moment is that the a slower card will limit a faster.

You also have issues with texture cache and a total lack of sharing geometry processing.

I'll say it again: if you can make a single card do the rendering for all of the subsamples, it'll be vastly more efficient.

Of course a single card will be more efficient, but on titles that are vastly CPU limited or have significant performance issues in multiple rendering due to RTT issues, whats the point of doubling the render power in those cases?

 

[Ailuros]

Of course a single card will be more efficient, but on titles that are vastly CPU limited or have significant performance issues in multiple rendering due to RTT issues, whats the point of doubling the render power in those cases?

I doubt a performance increase under current circumstances with RTT or other full screen effects is feasable; there might be a way but I'd figure it'd be quite complex to implement.

Definitely for those cases being able to scale AA sample density is a nice plus (if performance allows it of course). Highest EER bragging rights included

 

[KimB]

Of course a single card will be more efficient, but on titles that are vastly CPU limited or have significant performance issues in multiple rendering due to RTT issues, whats the point of doubling the render power in those cases?

Any program that has significant performance issues in rendering normally with multiple cards will also have significant performance issues with FSAA shared across the cards.

 

[Dave Baumann]

Any program that has significant performance issues in rendering normally with multiple cards will also have significant performance issues with FSAA shared across the cards.

The performance wouldn't be worse than a single card performance, you just get an increase in IQ (as opposed to nothing at all).

 

[KimB]

The performance wouldn't be worse than a single card performance, you just get an increase in IQ (as opposed to nothing at all).

Yes it would. Even in an ideal case the performance would be somewhat worse: you still need to combine the results from the two cards (which will take some memory bandwidth), and you still need to duplicate the data to send to both video cards. In cases that are hard on multi-card rendering, the situation will be much worse.

And then, if you decide to improve the texture LOD (if you don't, the textures will get blurrier, as they did with the Voodoo5), you'll have some major texture cache coherency problems.

Edit:
Oh, and finally, if ATI is choosing to do this, they'll really need to re-evalute their sample patterns. It would be rather silly to just use the sparse patterns they currently use on the separate cards.

 

[trinibwoy]

The performance wouldn't be worse than a single card performance, you just get an increase in IQ (as opposed to nothing at all).

Yes it would. Even in an ideal case the performance would be somewhat worse: you still need to combine the results from the two cards (which will take some memory bandwidth), and you still need to duplicate the data to send to both video cards.

And then, if you decide to improve the texture LOD (if you don't, the textures will get blurrier, as they did with the Voodoo5), you'll have some major texture cache coherency problems.

Chalnoth, ever consider that Dave has probably seen this thing in action and knows for a fact that performance is not lower in this mode?

 

[KimB]

Chalnoth, ever consider that Dave has probably seen this thing in action and knows for a fact that performance is not lower in this mode?

Well, that statement has about the same merit as:

"Chalnoth, ever consider that Dave has probably seen the pigs, and the fact that they can fly?"

And the simple answer is no, it's just not possible. Depending on the settings used, the performance hit may be small, but it is not going to be zero.

 

[trinibwoy]

Well, that statement has about the same merit as:

"Chalnoth, ever consider that Dave has probably seen the pigs, and the fact that they can fly?"

Ok boss.

 

[Dave Baumann]

Yes it would. Even in an ideal case the performance would be somewhat worse: you still need to combine the results from the two cards (which will take some memory bandwidth), and you still need to duplicate the data to send to both video cards. In cases that are hard on multi-card rendering, the situation will be much worse.

Combination isn’t necessarily and issue - should you not be gaining performance due to CPU then this is probably not going to be much of an effect given the bandwidth being talked about. Duplication of data will be an issue if you are interface limited, but that will limit all multiboard situations.

And then, if you decide to improve the texture LOD (if you don't, the textures will get blurrier, as they did with the Voodoo5), you'll have some major texture cache coherency problems.

That affects SSAA, but not MSAA.

And the simple answer is no, it's just not possible. Depending on the settings used, the performance hit may be small, but it is not going to be zero.

The likelihood of it occurring in real life is slim since we are talking about theoretical cases.

 

[KimB]

And then, if you decide to improve the texture LOD (if you don't, the textures will get blurrier, as they did with the Voodoo5), you'll have some major texture cache coherency problems.

That affects SSAA, but not MSAA.

But if you're using multiple cards for the effect, it makes no sense to use the same texture coordinates for both cards. Supersampling is essentially implied here (or, rather, a combination between the two).

 

[Dave Baumann]

Well, didn't we just provide a reason between us?

 

[KimB]

Well, didn't we just provide a reason between us?

No, because if you use the same texture LOD but use different texture coordinates, you will get similar performance to just one card doing the rendering. This will have the effect of reducing texture aliasing, but at the cost of some texture clarity. It's rather silly to waste that extra processing, if you ask me.

Regardless, even in this case, having the two cards themselves each use a mixed-mode FSAA rendering system and AFR or SFR (or supertiling) type rendering will allow for better texture LOD with a comparitively tiny performance hit, and the increased size of the z-buffer allows that to be compressed better, and there is the possibility of splitting the geometry load between the two cards, so no, it makes much more sense to not have the subsamples shared across multiple cards.

 

[Blazkowicz]

Any program that has significant performance issues in rendering normally with multiple cards will also have significant performance issues with FSAA shared across the cards.

The performance wouldn't be worse than a single card performance, you just get an increase in IQ (as opposed to nothing at all).

yes, a voodoo5 5500 doing FSAA 2x performs exactly like a voodoo4 doing no FSAA, and a voodoo5 6000 doing 4x performs like a voodoo5 5500 doing 2x.
even with render to textures/post-processing stuff unfriendly to NV SLI there still wouldn't be a performance loss, as each card does the same stuff on its own anyway?

 

[Dave Baumann]

having the two cards themselves each use a mixed-mode FSAA rendering system and AFR or SFR (or supertiling) type rendering will allow for better texture LOD with a comparitively tiny performance hit, and the increased size of the z-buffer allows that to be compressed better, and there is the possibility of splitting the geometry load between the two cards, so no, it makes much more sense to not have the subsamples shared across multiple cards.

You are not thinking along the right lines. Rules that exist on a single board don't change, think how and why those rules could possibly change when using two (hint: read back).

 

[Colourless]

yes, a voodoo5 5500 doing FSAA 2x performs exactly like a voodoo4 doing no FSAA, and a voodoo5 6000 doing 4x performs like a voodoo5 5500 doing 2x.
even with render to textures/post-processing stuff unfriendly to NV SLI there still wouldn't be a performance loss, as each card does the same stuff on its own anyway?

Yes, if it is setup properly. You send all the same data to both cards and just have differing sampling positions. Can even end up being a 'cheap' way to antialias render-to-texture targets (though there is a small chance that it might break with some titles that assume sampling positions with RTT).

Card A does the entire scene with one set of sampling positions.
Card B does the entire scene with a different set of sampling positions.

When buffer flip is done, Card B sends it's completed frame to Card A and Card A combines then sends them to scanout.

The only trouble that can come is when attmepting to lock the surfaces and read back from them. However the drivers should be able to handle that.

 

[KimB]

Yes, if it is setup properly. You send all the same data to both cards and just have differing sampling positions. Can even end up being a 'cheap' way to antialias render-to-texture targets (though there is a small chance that it might break with some titles that assume sampling positions with RTT).

Well, I suppose that is a good point. It will be more efficient at RTT, but you're still paying the cost of supersampling for something that we typically do with multisampling these days at a much lower performance hit. I still think that two cards + multisampling (with no shared FSAA) would give better performance, if an equivalent multisampling mode was available.