NDI Bandwidth & Latency: Capacity Planning Guide | Callaba
More live video workflow notes and product updates.
NDI bandwidth and latency depend on the NDI format, video format, sender, receiver and network path. “One NDI stream” is not a capacity unit. Plan from the exact source profile, measure real traffic, account for every receiver and preserve switch and endpoint headroom for production changes.
Capacity is consumed at the source, switch and receiver boundaries
Callaba provides NDI workflows, but not a guaranteed bandwidth or latency figure
Callaba includes NDI network configuration, discovery and adapters, and can use NDI in routing, monitoring and output paths. Operators can configure machine identity, discovery servers and explicit network addresses. Callaba does not currently document a built-in NDI capacity calculator, switch profiler or universal latency guarantee.
| Planning job | What Callaba covers | What to measure on site |
|---|---|---|
| Discover and address sources | NDI configuration, discovery and adapter workflows | Verify reachability and selected interface on the production network |
| Use NDI in a media path | Routing, Multiview and supported NDI output workflows | Measure actual sender egress, receiver ingress and switch load |
| Set a latency objective | Observe the managed workflow and downstream result | Measure glass-to-glass delay for the exact sender, receiver and display |
One planning view could bring together endpoint rates, selected adapters, switch telemetry and a timestamped source-to-output test. Until it exists, use Callaba for the NDI configuration and treat capacity and delay as properties of the measured site.
Identify NDI High Bandwidth, HX2 or HX3 first
The NDI families use different codecs and operating trade-offs. The current official NDI technical requirements identify SpeedHQ for NDI High Bandwidth and H.264 or H.265 for HX2 and HX3 certification. Compressed HX traffic can be much smaller, while encode and decode behaviour depends on implementation and settings. Do not shorten the older format name to “Full NDI”; the current official name is NDI High Bandwidth.
Ask both sender and receiver which format they actually use. A receiver that merely says “NDI compatible” may not accept every family, bit depth or alpha mode. Capture the negotiated result or observed stream details rather than relying on a product badge.
Use official figures as bounds, then measure the sender
The official NDI High Bandwidth table lists maximum examples for a defined SpeedHQ2 8-bit 4:2:2 profile: about 132.14 Mbps for 1080p60 and 249.99 Mbps for 2160p60. Those numbers are not universal rates for every NDI stream. Format, frame rate, codec profile, alpha, audio, metadata and implementation alter the result.
The official NDI HX2 format page gives approximate 1080p60 examples of 16 Mbps for H.264 and 11 Mbps for H.265. Treat those as format examples, not a promise for a device. HX2 and HX3 senders expose implementation and encoder choices, so record the configured or observed bitrate, GOP and codec rather than estimating from the HX name alone.
Measure a representative programme for several minutes, including high motion, graphics and the actual audio layout. Record average, high-percentile and short peaks on the sender port. Plan from the observed profile plus an explicit operational reserve chosen by the network owner, not from a calculator with missing inputs.
Count copies according to unicast or multicast behaviour
With unicast, one sender may transmit a separate stream to each receiver. Three receivers can therefore create roughly three media copies on the sender's access link, subject to the actual implementation. The official NDI multicast guidance describes multicast UDP media with forward error correction, while each receiver still opens a unicast TCP connection for bidirectional metadata such as tally or PTZ.
Multicast requires IGMP, a functioning querier and snooping so subscribed ports receive the media instead of every switch port. Draw the receiver count and physical path for preview, programme, Multiview, recording, graphics and confidence monitoring. Confirm destination addresses, group joins and leaves, querier state and residual unicast metadata traffic; discovery alone does not prove multicast filtering.
Work a port and uplink budget from simultaneous streams
Suppose one 1080p60 NDI High Bandwidth source is measured at the official 132.14 Mbps maximum example and four unicast receivers run simultaneously. The media-copy budget is 4 × 132.14 = 528.56 Mbps on the sender port. If measured control, metadata and other traffic peaks at 70 Mbps and the network owner reserves 150 Mbps for operational change, the documented planning load becomes 748.56 Mbps. Compare that value with the configured usable port capacity and measured packet behaviour; do not turn the example's reserve into a universal percentage.
Repeat the arithmetic at every converged uplink. Eight distinct 132.14 Mbps source copies would total about 1,057.12 Mbps before control traffic or reserve, so a one-gigabit uplink cannot be accepted on arithmetic alone. Use observed simultaneous copies, not the total number of source names, and re-run the calculation for failover or rehearsal states that add receivers.
Check every oversubscription point, not only the desktop port
A one-gigabit endpoint can appear comfortable while several sources converge on an uplink. Sum the measured high-percentile traffic for streams that can coexist on each access port, uplink and switch fabric. Include both directions where devices send and receive, plus ordinary control, storage and management traffic.
Read interface utilization together with drops, errors, queue discards and pause behaviour. Average utilization can hide microbursts. If drops begin during source cuts or receiver joins, investigate the affected egress queue and topology before blaming the codec.
Latency has capture, codec, network, receive and display components
Official NDI certification requirements list defined test targets: less than 100 ms for NDI High Bandwidth, 100 ms for HX3 and 150 ms for HX2 in the stated certification contexts. Those are product-certification criteria, not a guarantee for the deployed production path. Camera processing, frame synchronization, encoding, network queues, decoding, application buffers and display refresh all contribute.
Measure glass-to-glass with a visual time source or another repeatable marker visible at capture and output. Run enough samples to report a distribution, not one photograph. In parallel, capture network transit and endpoint queue evidence so the team can locate a regression.
Discovery and media reachability are separate checks
A source can appear in a discovery list while its media path uses the wrong interface or cannot reach the receiver. Conversely, explicit discovery-server or address configuration may make a routed source usable without broad multicast discovery. Document which method the production uses and which subnets and adapters are intended.
Pin or prioritize the correct interface where supported, remove ambiguous routes and verify bidirectional control reachability. Then receive the actual media and inspect rate, frames and audio. A discovered source name is not a completed media test.
Example: sources are visible, but Multiview stutters during rehearsals
All sources remain discoverable and individual point-to-point tests look clean, yet the confidence wall drops frames when several receivers open. Competing hypotheses are decoder saturation, sender overload, discovery churn or a congested switch uplink. The first comparison aligns sender egress, receiver ingress, decoder load and switch queue drops during a controlled receiver join. Sender rates and decoder load stay stable while uplink egress discards rise with every added unicast copy. The root cause is the shared network boundary. Reduce duplicate flows or redesign the topology, then repeat the join test and confirm both zero new drops and smooth displayed motion.
Build a measured NDI acceptance record
- Inventory streams. Record family, video format, audio, sender, receivers and intended adapter.
- Measure one source. Capture average, high-percentile and peak rates with real programme content.
- Expand receiver count. Observe sender egress and multicast state as each consumer joins.
- Inspect the topology. Check access ports, uplinks and fabric for load, errors, drops and queues.
- Measure displayed delay. Repeat glass-to-glass samples in ordinary and high-load states.
- Exercise recovery. Restart a source, receiver and switch path separately; confirm discovery and media return.
NDI bandwidth and latency FAQ
How much bandwidth does a 1080p60 NDI stream use?
It depends on the NDI family and implementation. The official SpeedHQ2 example lists about 132.14 Mbps maximum for NDI High Bandwidth; measure the actual sender.
Does each NDI receiver multiply bandwidth?
It can with unicast. Multicast may avoid repeated sender copies, but requires a correctly designed and observed multicast network.
Is NDI codec latency the same as glass-to-glass latency?
No. Capture, buffering, network, decode and display add delay beyond the codec component.
Does Callaba guarantee an NDI bandwidth or latency?
No. Callaba provides NDI configuration and media workflows, but capacity and delay still have to be measured on the deployed path.