I know this question sounds a little confusing now, but I mean SLI in its original meaning - scanline interleave.
Well, SLI already has returned - or rather, is announced to return. Today it just carries the name SuperAA. Strange, eh?
Generally, you can use multiple GPUs to either increase the framerate at a given resolution, or to increase the resolution while keeping the framerate the same.
When AFR or SFR (be it a single split or tiles) work, all is well and you either get a massive increase in fps or you can more than double the resolution (efficiency increases with resolution) while getting the same framerate.
But AFR breaks with framebuffer locks, and textures rendered to in the previous frame need to be transferred or re-rendered. With SFR, RTT textures generally need to be transferred or rendered twice.
Rendering textures twice does of course impact the fps you can reach, but even if each card rendered every texture, you still would be able to exactly double the final output resolution at the same framerate, ignoring the overhead. Well, but only if the resolution of those textures does not scale with output resolution.
Post processing effects, however, as well as a few other things, do scale with output resolution. So if you rendered the whole scene into a framebuffer-sized texture, then perform a simple bloom filter on it, your SFR system, set up to re-render textures on each card, would see close to no speed-up at all. And if that texture is used in the next frame as well, e.g. for light trails, AFR will significantly slow down, too.
Now this is where SuperAA kicks in. A simple mode that pretty much guarantees higher AA levels at the same framerate. Because the output resolution stays the same. But does it really? Isn't supersampling just a higher resolution downfiltered?
With SuperAA, each card renders as if it were the only card in the system, bar a possible LOD bias. Each card renders its own RTT textures. No communication until scanout.
This is exactly how scanline interleave works, too, except that at scanout the lines are interleaved instead of blended. And that the application is aware of the increased output resolution.
What if we took this awareness away?
Say, you have a game that delivers playable framerates at 1024x768 and whatever settings you like with a single card setup, but not above. Unfortunately, AFR and SFR both do not work as expected, but you have the option to enable SuperAA on a dual-card machine. So you get nice smooth edges at the same playable framerate. Unfortunately, 1024x768 is a bit low for a high quality monitor. Wouldn't it be nice, then, to be able to just double the resolution to 1440x1080 instead of having 2xSSAA?
Well, SLI already has returned - or rather, is announced to return. Today it just carries the name SuperAA. Strange, eh?
Generally, you can use multiple GPUs to either increase the framerate at a given resolution, or to increase the resolution while keeping the framerate the same.
When AFR or SFR (be it a single split or tiles) work, all is well and you either get a massive increase in fps or you can more than double the resolution (efficiency increases with resolution) while getting the same framerate.
But AFR breaks with framebuffer locks, and textures rendered to in the previous frame need to be transferred or re-rendered. With SFR, RTT textures generally need to be transferred or rendered twice.
Rendering textures twice does of course impact the fps you can reach, but even if each card rendered every texture, you still would be able to exactly double the final output resolution at the same framerate, ignoring the overhead. Well, but only if the resolution of those textures does not scale with output resolution.
Post processing effects, however, as well as a few other things, do scale with output resolution. So if you rendered the whole scene into a framebuffer-sized texture, then perform a simple bloom filter on it, your SFR system, set up to re-render textures on each card, would see close to no speed-up at all. And if that texture is used in the next frame as well, e.g. for light trails, AFR will significantly slow down, too.
Now this is where SuperAA kicks in. A simple mode that pretty much guarantees higher AA levels at the same framerate. Because the output resolution stays the same. But does it really? Isn't supersampling just a higher resolution downfiltered?
With SuperAA, each card renders as if it were the only card in the system, bar a possible LOD bias. Each card renders its own RTT textures. No communication until scanout.
This is exactly how scanline interleave works, too, except that at scanout the lines are interleaved instead of blended. And that the application is aware of the increased output resolution.
What if we took this awareness away?
Say, you have a game that delivers playable framerates at 1024x768 and whatever settings you like with a single card setup, but not above. Unfortunately, AFR and SFR both do not work as expected, but you have the option to enable SuperAA on a dual-card machine. So you get nice smooth edges at the same playable framerate. Unfortunately, 1024x768 is a bit low for a high quality monitor. Wouldn't it be nice, then, to be able to just double the resolution to 1440x1080 instead of having 2xSSAA?
