Rasterization is the process of converting vector-based geometric objects into pixels on a screen. While you could simply check the center of a pixel to see if a polygon covers it (point sampling), conservative rasterization tests coverage at the corners of the pixel leading to much more accurate results. The Intel® Processor Graphics Gen9 GPU architecture adds hardware support for conservative rasterization. There is a flag in the shader to indicate whether the pixel is fully (inner conservative) or partially covered (outer conservative).
The implementation meets the requirements of tier2 hardware per the Direct3D specification. It is truly conservative with respect to floating point inputs, and is at most 1/256th of a pixel over conservative (tier 2). No covered pixels are missed or incorrectly flagged as fully covered. Post-snapped degenerate triangles are not culled. A depth coverage flag notes whether each sample was covered by rasterization and has also passed the early depth flag test (see SV_DepthCoverage in Direct3D).
Outer Conservative Inner Conservative
Figure 2. Conservative rasterization.