G-buffer
The set of render targets deferred shading fills in its first pass — a per-pixel description of the surface before any light is computed.
The name is short for "geometry buffer" and was coined by Saito and Takahashi in a 1990 paper on comprehensible rendering. The buffer stores what lighting needs to know about every visible pixel: depth, normal, albedo, and material parameters such as roughness and metalness. The second pass reads it and computes light once per pixel, no matter how many objects overlapped there.
The design choice is a trade-off between "fat" and "thin" — more channels give richer materials but devour bandwidth, which is the deferred approach's Achilles heel. Normals are often stored compressed to save room, and transparency does not fit at all, because each pixel can hold only a single surface.