AMD DX11 Leo Demo

lollerwaffle said:
Ran fine on my 5850/2600k.
Funny, ran like crap on my 6750.

Ashbringer, are you serious?
Let's see 6750 = 5750 core with the addition of hdmi 1.4 and UVD3.
http://www.overclockers.com/sapphire-hd-6750

.sarcasm on.
I wonder why it sucked on your low end <"$100 when released last year" video card when the demo was written for a $550 just release top-of-the-line 7970.
.sarcasm off.


Since when do game developers need these demos? I'm sure they have all kinds of whitepapers to read and their own demos they can develop and test things out... they don't need AMD to produce a demo to show them anything.

I'm a software Engineer and would much rather a dry technical explanation come with a demo. I don't *NEED* a demo, but it certainly helps, and never hurts.

If the vendor provides me sample code to tweak, rather than having me write it from scratch to test out new features, it's a plus cause it saves me time and effort.
 
and furthermore... yes it was cute... But I don't want cute. I want stuff blowing up; fire and explosions... I want awesome, not cute.
 
I don't understand exactly what it was showcasing, something with the lighting. Anyways I like tech demo's and CG, I am impressed for something to look that nice in-engine, and quite enjoyed it.

Well, there are two main methods to render 3d games. Forward Rendering was the original, old school one; Deferred Rendering was the new kid on the block. Each one had it's pros and cons, and the main pro of the deferred rendering was that it had a much lesser performance hit when dealing with multiple light sources and was a relatively simple method; because of that Deferred Rendering became eventually the "go to" method for most important rendering engines.

The problem is that D.R. brings a list of cons in the form of heavier performance hit when handling multiple materials, and because it usually discards the geometry data it can't really apply proper MultiSampling antialiasing. The solution for 1) chosen by most devs was "ok, we won't use multiple materials! i mean, texturing alone is good enough, right?" the solution for 2) was "we will create a special buffer, let's call it G buffer, and we will store geometry data and some other useful things there!". Problem is that the G buffer eats memory like nothing, and if you try to use different materials, each material makes the G buffer even bigger (on top of the performance hit).

At AMD this D.R. became a royal pain in the ass, since their gpu designs were made with proper MSAA in mind, and thus this workaround has made them take significantly bigger performance hits than on Nvidia's hardware, and thus they have been hard at work until they got this solution.

The solution they found is to run a proper compute shader to apply the lighting to the Forward Rendered image, instead of the usual way of "Render everything 1 time for each light source in the scene!" This way they save a great deal of passes, they save on memory by not needing the G buffer (the geometry is always present on a Forward Renderer, instead of discarded), they can use the proper MSAA included on their original gpu design, and multiple materials can be used without the big performance & memory hit of the DR, all it takes is compute time for the new shader.

So, this demo to a game dev should be interesting in the form of the performance achieved with a FR engine while dealing with multiple light sources, the lack of a noticeable performance and memory useage hit due to the multiple materials & the use of "proper" MSAA.


Let me know if i missed something.
 
Well, there are two main methods to render 3d games. Forward Rendering was the original, old school one; Deferred Rendering was the new kid on the block. Each one had it's pros and cons, and the main pro of the deferred rendering was that it had a much lesser performance hit when dealing with multiple light sources and was a relatively simple method; because of that Deferred Rendering became eventually the "go to" method for most important rendering engines.

The problem is that D.R. brings a list of cons in the form of heavier performance hit when handling multiple materials, and because it usually discards the geometry data it can't really apply proper MultiSampling antialiasing. The solution for 1) chosen by most devs was "ok, we won't use multiple materials! i mean, texturing alone is good enough, right?" the solution for 2) was "we will create a special buffer, let's call it G buffer, and we will store geometry data and some other useful things there!". Problem is that the G buffer eats memory like nothing, and if you try to use different materials, each material makes the G buffer even bigger (on top of the performance hit).

At AMD this D.R. became a royal pain in the ass, since their gpu designs were made with proper MSAA in mind, and thus this workaround has made them take significantly bigger performance hits than on Nvidia's hardware, and thus they have been hard at work until they got this solution.

The solution they found is to run a proper compute shader to apply the lighting to the Forward Rendered image, instead of the usual way of "Render everything 1 time for each light source in the scene!" This way they save a great deal of passes, they save on memory by not needing the G buffer (the geometry is always present on a Forward Renderer, instead of discarded), they can use the proper MSAA included on their original gpu design, and multiple materials can be used without the big performance & memory hit of the DR, all it takes is compute time for the new shader.

So, this demo to a game dev should be interesting in the form of the performance achieved with a FR engine while dealing with multiple light sources, the lack of a noticeable performance and memory useage hit due to the multiple materials & the use of "proper" MSAA.


Let me know if i missed something.

You are kind of overstating the cons of deferred renderers I think. Deferred lighting (not to be confused with shading) tries to remedy the heavy G-Buffer cost and material problems by having another geometry pass... light setup may be slightly cheaper and there's less bandwidth hit, but at this point GPU bandwidth (in PC land) is so fucking enormous, that for any geometry heavy scene at all, it will probably lose. On PC, memory and bandwidth is probably the least of your worries on anything that isn't a pile of shit.

There's an overhead to render setup in deferred cases, but you have other opportunities like better batching, cleaner pipeline, potentially more predictable performance, etc. It can be a win, maybe a lose, but in the direction most games seem to be going, it seems to err on the side of "good idea" more often than not, it enables lots of flexibility in its own way. The material "problem" has been worked around tons of times already (Little Big Planet 2 comes to mind, with a similar kind of look to this demo, and it's deferred) so it's by no means an impossibility. It's not like "welp the entire game world has to look like metal we're fucked."

MSAA has worked properly in deferred renderers since DX10, and in DX11 has nicer performance characteristics yet. Though deferred or not, MSAA doesn't scale nicely and shader based AA looks like the future for numerous reasons. Compute shader already seems to lend itself well to deferred renderering, as evidenced by BF3, further helping its case with overdraw and bandwidth.
 
Back
Top