Game Development Reference
Forward (with transparency)
Forward (no transparency)
Deferred (no transparency)
(a) Render time
(b) Memory usage
Figure 4.4. (a) Render time and (b) memory usage for tiled forward and tiled deferred
shading with varying MSAA settings. We were unable to allocate the 8 × MSAA frame-
buffer for deferred, which is why no timing results are available for that configuration.
Memory usage estimates are based on a G-buffer with 32-bit depth, 32-bit ambient, and
64-bit normal, diffuse, and specular components (using the RGBA16F format).
One interesting note is that our unoptimized Z-reduction scales almost linearly
with the number of samples: from 0 . 75 ms when using one sample to 5 . 1mswith
MSAA. At that point, the contribution of the Z-reduction is quite significant
with respect to the total frame time. However, it still provides a speedup in
our tests. It is also likely possible to optimize the Z-reduction step further, for
instance, by using compute shaders instead of a fragment shader.