Why AV1 Low Latency Omits Bidirectional Prediction

In ultra-low-latency (ULL) video streaming, minimizing the glass-to-glass delay between video capture and display is the primary engineering objective. While the AV1 video codec incorporates sophisticated bidirectional prediction to maximize compression efficiency, this technique is intentionally omitted in real-time configurations. Enabling bidirectional prediction introduces unavoidable buffering, frame reordering, and computational complexity that directly violate the tight timing budgets required for interactive workflows like cloud gaming, video conferencing, and live interactive broadcasting.

How Bidirectional Prediction Works

In standard video compression, predictive coding reduces temporal redundancy by referencing other frames. While unidirectional prediction (P-frames) only references previously displayed frames, bidirectional prediction (B-frames or B-pictures) allows a frame to reference both past and future frames in the display order. AV1 takes this further with advanced compound prediction modes, allowing motion vectors to combine multiple temporal references to achieve superior image quality at lower bitrates.

The Pipeline Reordering Delay

The primary reason bidirectional prediction is disabled in ULL environments is the requirement for frame reordering.

For an encoder to predict a current frame using a future frame, that future frame must be captured, processed, and transmitted first. This creates an inherent structural delay:

Capture Delay: The encoder must wait for future frames to arrive in the buffer before it can begin compressing the current frame.
Decode Delay: The decoder cannot present the current frame until it has received and decoded the subsequent reference frame.

In interactive video workflows operating via protocols like WebRTC, total glass-to-glass latency must often remain below 100 to 150 milliseconds. Introducing even two or three frames of reordering delay at 60 frames per second adds 33 to 50 milliseconds of pure algorithmic latency, consuming a massive fraction of the allowable latency budget before network transmission even occurs.

Encoder Computational Overhead

AV1 is computationally demanding. Evaluating bidirectional motion vectors requires searching reference spaces across multiple directions (past and future) and evaluating complex compound weighting modes.

In ULL workflows, the encoder must operate in real time without spikes in frame processing time. Omission of bidirectional prediction significantly reduces the motion estimation search space. This allows software and hardware encoders (such as SVT-AV1 in low-latency modes or dedicated NVENC/VCN hardware) to process and packetize frames immediately upon capture, ensuring deterministic per-frame encoding times.

Network Resiliency and Packet Loss Recovery

Real-time streaming typically relies on unreliable transport protocols (UDP) where packet loss is frequent. Bidirectional reference structures create complex dependency chains. If a reference frame positioned in the future is lost or corrupted in transit, all intervening bidirectionally predicted frames become undecodable.

By restricting AV1 to forward-only prediction chains (an IPPP... structure), error recovery becomes much simpler:

Reference Picture Invalidation: The decoder can signal loss via RTCP feedback, and the encoder can immediately reference an older, acknowledged frame.
Long-Term Reference (LTR) Frames: Encoders can fall back on known valid states without clearing large reorder buffers.
Gradual Refresh: Techniques like intra-refresh (refreshing columns of macroblocks across P-frames) function seamlessly without out-of-order frame dependencies.

The Standard ULL Configuration for AV1

When tuning AV1 for ultra-low latency, engineers configure the codec for a "zero-latency" or low-delay tuning. This forces the encoder to operate strictly with a lookahead buffer size of zero and disables all non-causal reference tools. The modest loss in compression efficiency is offset by an immediate reduction in buffering delay, predictable compute performance, and robust recovery in variable network conditions.