UL 3DMark gets a new feature test - Mesh Shader

neeyik

Posts: 2,963   +3,628
Staff member
UL have updated 3DMark to include a new feature test, that examines the potential benefits of using mesh shaders. This a geometry pipeline system in Direct3D 12 and has been around for a little while now, although Microsoft have been promoting it more since the release of DirectX 12 Ultimate.

Without going into the boring details, mesh shaders basically allow developers to have a lot more control over how geometry is processed in the rendering cycle. Probably the biggest benefit comes from being able to work on small portions of a mesh (a collection of primitives that make up an entire model) rather than how it normally is done -- I.e. process the whole mesh first and then decide which bits of it are going to be visible and therefore required for the frame.

noMS.jpg

3DMark's Mesh Shader test runs through a simple scene, containing a crap ton of geometry, in two modes: first is a fairly normal method, bar it's using compute shaders to determine the level of detail for a mesh and what parts are not visible (and therefore can be culled). This is what's shown in the above image.

MSused.jpg

The second mode using the mesh shader pipeline, which figures out which parts of a mesh aren't visible, and discards them -- only the visible ones are passed on for further processing.

This is what's going on in the second image. It obviously looks almost exactly the same, however notice the better lighting of the statue's plinths, especially the corners. Possibly a minor glitch in the test, but could also be down to better select of LOD, when using the mesh shader pipeline.

meshlets.jpg

Here you can see all of the sub-sections of each mesh (called meshlets) that are checked for visibility, before being culled.

All of this is very nice, but what about the actual result? For my PC, a 4.5x increase in the test frame rate:


Naturally, the use of mesh shaders isn't an instantaneous performance boost to a game that's using lots of geometry, as the performance bottleneck may not be in the geometry section of the rendering. But any performance gains that are essentially free, if one has the programming know-how and the user has the right graphics cards, isn't to be sniffed at.
 
Back