Low-latency compression: choose between bit-exact, visually lossless and video coding

First identify what the compression is for

Compression between internal processing blocks, across a display link and into a stored video stream solves different problems. Draw the insertion point, data format and receiving endpoint before choosing a codec. A proprietary framebuffer format may need a matching decompressor; a standards-based video stream must match the destination decoder configuration.

VESA describes DSC as low-latency, visually lossless compression for video interfaces. Visually lossless means a perceptual quality objective; it does not mean every reconstructed sample equals its input. If measurement, further processing or exact reconstruction requires unchanged samples, require bit-exact losslessness and test that property explicitly.

Sources: 1

Calculate the uncompressed baseline first

For illustrative UHD 3840 × 2160 at 60 fps, tightly packed 8-bit YCbCr 4:4:4 uses 24 bits/pixel and about 1.493 GB/s of active payload. At 4:2:0, the average is 12 bits/pixel, about 0.746 GB/s. These are ideal active-pixel figures; packing, alignment and transport overhead may increase the actual traffic. Changing chroma sampling changes the input representation and must not be counted as a codec-only gain.

If a vendor quotes an average 2:1 lossless ratio, test high-entropy inputs and ask how expansion is handled. Lossless compression cannot guarantee that every possible input becomes smaller. Provision fallback, metadata and buffer space using worst-case behavior. For a fixed-rate design, validate the chosen quality target at that rate with the actual content set.

Define the latency measurement boundary

At 60 fps, a frame interval is 16.67 ms. A quoted sub-millisecond codec delay therefore needs a precise definition: first pixel to first output, slice availability, complete-frame completion, or encode-plus-decode processing. It does not automatically establish camera-to-display latency. Include sensor exposure/readout, queues, transport, decoding and display behavior in the end-to-end budget.

Ask how slice or line sizes affect buffering, parallelism, quality and endpoint interoperability. Test backpressure and recovery when the downstream consumer stalls. A very small measured core delay is valuable only if surrounding buffers and the required picture structure preserve that advantage.

Use an acceptance matrix for the actual link

Compare reconstruction requirement, interoperability, bitrate behavior, total delay, memory traffic and implementation cost. For visually lossless operation, assess difficult textures and repeated encode/decode stages if they occur in the product. For bit-exact operation, compare every reconstructed sample. Select a method for the specific data boundary instead of ranking all compression IP by one advertised ratio.

Bring these to your IP evaluation

• Specify exact versus perceptual reconstruction requirements.
• Budget worst-case traffic, expansion and fallback buffers.
• State core latency and end-to-end latency separately.