Regarding PhysX on CPU:
Oles Shishkovtsov: We conducted our AAA along with FXAA to achieve more control per pixel in order to adjust the blur and sharpness better. There is now no way to disable AAA, this is due to the low implementation in the rendering pipeline. The side effect is, of course, that the image is always smoothed to a certain degree. The usual problem of AAA remains, however, a lack of sub-pixel information to shader fragment or small pixel capture.
The solution is to give the player the option to hand to activate certain SSAA modes. You'd be surprised how little additional information is necessary in order to achieve the quality of 4x MSAA or to offer! Basically provides 2x SSAA (1.41421 x 1.41421 per axis) together with FXAA, which is applied in the supersampling resolution to hit enough information to 4x MSAA in all respects, even in terms of performance - with a deferred renderer . On top of the sub-pixel aliasing is greatly reduced with shaders.
For those who are interested: An Ordered or rotated grid is great for all vertices which occur in the pixel grid when the box filter is used. If you use something more sophisticated as we (FXAA), then the filter is no longer able to cover the grid in the correct orientation - which means a substantial loss of quality. Although a slight rotation of the image from a qualitative point of view is very good, it takes a lot of power and storage space. So be it, traditional MSAA is dead - Edge smoothing is the responsibility of the programmer and not a magical hardware feature.
Oles Shishkovtsov: Absolutely, but an extremely fast CPU is needed to catch up with GPUs. Some things are obscenely well parallelized and thus it is difficult to achieve the performance of a 2.0 TF graphics card with a 0.1 TF processor.
Oles Shishkovtsov: In short, everything. The long answer is, there are two selectable modes: one tessellated only by our artists selected objects and the other almost everything. The power loss is around 15 percent, respectively three, mind depending on the GPU architecture.
Oles Shishkovtsov: No, we do not deal with it. Why you should play with DX9, DX11 when the same quality is 10 to 15 percent faster? DX9 is contained in Last Light for compatibility reasons only.
Oles Shishkovtsov: Apart from tessellation and parallax occlusion mapping corresponds to the DX11 path largely the DX10 version as we achieve more than 95 percent of the PCs in our target group. Purely from the performance point of DX11 is the fastest, followed by DX10 and DX9 [Anm.dR: without tessellation and POM].
I think we played a great job, especially in scalability. For example, Metro Last Light is running in DX11 mode on an Intel HD Graphics 4000 is passable - with small details of course - and even Haswell pretty good. Nevertheless, we force with high 4x super-sampling anti-aliasing and very high details and a Geforce GTX 680 or Radeon HD7970 in the knee.