Choosing forward, deferred, or Forward+ rendering

Start with the workload

A rendering architecture determines when visible surfaces receive lighting and which intermediate data must survive. Choose it against representative scenes: a small interior, a crowded night street, and a view dominated by glass can stress different resources. Count overlapping lights and shaded pixels, not only lights placed in the level.

Separate the three mechanisms

Forward rendering evaluates material and lighting together. Deferred shading first stores opaque surface attributes in a G-buffer, then lights those attributes; extra storage and material representation become architectural concerns. Transparency commonly needs a separate path, and multisample antialiasing (MSAA) requires additional care. [1][3]

Forward+ retains forward material shading but constructs spatial light lists, often using compute-generated screen tiles. Fragments then evaluate candidate lights from their tile. AMD's sample provides an original implementation of this division of work. [2]

Work through a bandwidth estimate

Consider an illustrative 1920 × 1080 G-buffer storing 16 bytes per pixel. One complete image occupies 33,177,600 bytes, about 31.6 MiB, before depth, multisampling, or auxiliary buffers. One full write and one full read at 60 frames per second represent about 3.98 GB/s of logical traffic.

This arithmetic is a planning estimate, not measured external-memory bandwidth: compression, caches, tile-local storage, and extra passes change actual traffic. Use it to ask whether an additional material channel justifies its cost. For Forward+, make a corresponding estimate for tile headers, light indices, and overflow handling.

Run a controlled comparison

Keep resolution, lights, materials, shadow quality, and exposure identical. Capture GPU timings for depth preparation, list construction, opaque shading, transparency, and resolves. Record list-length distributions as well as the worst tile; an average can hide a bright intersection.

Include a camera moving through particles and glass. If lists use opaque depth bounds, explicitly verify that transparent fragments retain every relevant light. Visualize candidate counts and force the capacity limit to confirm that overflow produces a defined fallback rather than missing illumination.

Recognize the boundaries

A light list reduces irrelevant lighting work; it does not automatically reduce shadow rendering, material complexity, or transparency overdraw. Deferred MSAA is possible, but preserving and shading sample information has costs. [1][3]

Treat the choice as a measured budget decision. A hybrid can be appropriate, but every additional path increases shader maintenance and image-matching work. Document the scenes that justified the decision and repeat the comparison when the content mix changes.

Sources

• [1] Forward, Forward+, and Deferred — Khronos Group.
• [2] ForwardPlus11 sample — AMD GPUOpen.
• [3] Deferred Shading in Tabula Rasa — Rusty Koonce, NVIDIA GPU Gems 3.