HEVC Architecture Basics
HEVC keeps the prediction-and-residual philosophy of H.264 but replaces the fixed macroblock grid with nested quadtrees. Blocks can be as large as sixty-four samples or as small as eight, and the encoder chooses the partition that best matches local content. This single change is responsible for much of the efficiency gain.
Coding tree units and their parts
A coding tree unit decomposes recursively into coding units. Each unit carries its own prediction decisions, and prediction can be applied in units of different granularity from the coding unit itself. That decoupling lets the codec predict at a scale suited to the texture while coding at a scale suited to the statistics.
Prediction and transform improvements
Intra prediction gains more directional modes, including planar modes that model smooth gradients. Inter prediction gains a merge mechanism that signals motion by reference to a neighbouring candidate rather than coding a vector afresh, plus a wider range of interpolated fractional positions for sub-pixel motion.
On the transform side the block sizes grow, which helps stationary areas, and residual coding gains improved context modelling. Sampling formats such as 4:2:0 in ten bits become common because the extra precision suppresses banding on gradients without a large rate penalty.
Practical consequences
| Attribute | Observation | Planning effect |
|---|---|---|
| Partitioning | Content-adaptive | Large gain on mixed detail |
| Motion signalling | Merge-based | Cheaper at high motion |
| Bit depth | Ten-bit common | Smoother gradients |
| Complexity | Higher than H.264 | Slower encodes, more parallel gain |
Continue to the AV1 page for the tools that go beyond this design, or revisit H.264 for the baseline.