Best AV1 Presets for Low-Latency Live Streaming

AV1 delivers exceptional compression efficiency, but its computational complexity has traditionally made real-time processing difficult. Modern software encoders like SVT-AV1, libaom, and rav1e, alongside dedicated hardware encoders, now offer optimized speed presets specifically designed for real-time, low-latency transmission. Achieving low latency requires balancing encoding speed against compression efficiency to prevent frame drops and pipeline buffering.

SVT-AV1 Presets

SVT-AV1 is the most widely adopted open-source AV1 encoder for high-throughput production. Its speed presets range from 0 (slowest, highest quality) to 13 (fastest).

• Preset 10 to 13: These presets are mandatory for real-time live streaming on standard modern multi-core CPUs (e.g., 8 to 16 cores). They strip away heavy compute passes, like deep visual metrics optimization and deep partition searches, ensuring encoding speeds keep pace with 1080p60 or 4K30 incoming video feeds.
• Preset 8 and 9: These are usable on high-end, high-core-count workstation or server processors (such as AMD Threadripper or EPYC). They offer significantly better quality-per-bit than presets 10–13, but they demand substantial processing power to avoid falling behind real-time playback.

To enforce low latency in SVT-AV1 alongside these presets, you must also disable frame lookaheads (--lookahead 0) and enable low-delay mode (such as --pred-struct 1 for a flat prediction structure).

libaom-AV1 (Real-Time Mode)

The reference encoder libaom provides a dedicated real-time mode (--rt) combined with the --cpu-used switch, which controls speed from 0 to 10.

• --cpu-used 7 to 10: These presets are engineered specifically for real-time communication protocols like WebRTC. In --rt mode, levels 7 through 10 skip exhaustive search algorithms, allow non-reference frame dropping if the pipeline stalls, and implement single-pass encoding. Level 8 is the industry standard for software-based WebRTC implementations, while levels 9 and 10 are ideal for resource-constrained systems.
• Levels below 6 in libaom are unsuitable for live streaming, as they introduce severe latency and cannot maintain a 30fps or 60fps encoding rate in real time.

rav1e Presets

rav1e, the memory-safe Rust-based AV1 encoder, uses speed levels from 0 to 10.

• Speed 10: This is the only preset capable of sustaining live, real-time encoding for HD video on standard consumer hardware. It disables multi-pass decisions, complex transform units, and extensive RDO (Rate-Distortion Optimization).
• Speed 8 and 9: These can achieve real-time speeds on dedicated server hardware with lower-resolution streams (such as 720p), but they often struggle to maintain real-time low-latency compliance on standard client machines.

Hardware Encoders (NVENC, Quick Sync, AMF)

For production environments where sub-second latency is critical, dedicated hardware silicon bypasses CPU bottlenecks entirely:

• NVIDIA NVENC (RTX 40-series and newer): Presets P1 (Fastest) through P4 (Medium) provide reliable sub-frame encoding times. For ultra-low latency, P1 to P3 are ideal because they minimize internal buffering while still outperforming older h.264 codecs in visual fidelity.
• Intel Quick Sync Video (Arc and 11th Gen Core or newer): The Very Fast and Faster target usage presets eliminate latency overhead and provide stable real-time throughput up to 4K60.
• AMD AMF (Radeon RX 7000-series): The Speed preset configures the hardware VCN block for low-delay mode, disabling lookahead buffers.

Essential Low-Latency Configuration Rules

Regardless of the selected preset, the encoder will fail to deliver low latency unless associated pipeline settings are adjusted:

1. Disable B-Frames: B-frames require future frame references, inherently introducing presentation delay. Set maximum consecutive B-frames to 0.
2. Zero Lookahead: Disable multi-frame analysis buffers (set lookahead to 0 or 1).
3. Strict Constant Bitrate (CBR): Use CBR with minimal VBV (Video Buffer Verifier) buffer sizes to prevent large frame spikes that cause network transmission delays.