Network Device Interface

NDI is designed to run over gigabit Ethernet,[1] with the NDI codec[2] expected to deliver 1080i HD video at VBR data rates typically around 100 Mbit/s.[3]

By Default, NDI uses the mDNS (Bonjour / Zeroconf) discovery mechanism to advertise sources on a local area network, such that NDI receiving devices can automatically discover and offer those sources, although 2 other discovery modes (NDI Access, NDI Discovery Server) allow for operations across subnets, and without mDNS. Sources are created using an arbitrarily selected TCP port from a range of ports on the NDI send host. When a source is requested, a TCP connection is established on the appropriate port with the NDI receiver connecting to the NDI sender. NDI 3.x has options to use UDP multicast or unicast with forward error correction (FEC) instead of TCP, and can load balance streams across multiple network interface controllers NICs without using link aggregation. NDI 4.0 introduces multi-TCP connections.

NDI carries video, multichannel uncompressed audio and metadata. Metadata messages can be sent in both directions allowing the sender and receiver to message one another over the connection with arbitrary metadata in XML form.[4] This directional metadata system allows for functionality such as active tally information fed back to sources to understand that they are on-air (program / preview). NDI also allows senders to determine the number of connected receivers, so they can skip unnecessary processing and network bandwidth utilisation when there are no NDI receiver clients connected. NDI Receivers can opt to connect to various combinations of streams, to support things like audio-only or metadata-only connections where video is not required.

The NDI software development kit (SDK) is available for Windows, Linux and macOS,[5] and has also been ported to iOS, tvOS,[6] Android, Raspberry Pi, and FPGA.[7][8] The SDK is available via a royalty-free proprietary license.[9]

Other IP video protocols emerging for use in professional video production (rather than IP video used for distribution to end users) include SMPTE 2022, SMPTE 2110, ASPEN (largely superseded by SMPTE 2110) and Sony NMI. There are clear differences in the technology used by these protocols.

NDI was publicly revealed by NewTek on 8 September 2015 and was demonstrated at the IBC broadcast exhibition in Amsterdam that week.[17] The first device shown using NDI was the NewTek TriCaster which delivered an NDI feed from each of its SDI inputs as well as four output feeds from its vision mixer. The TriCaster could also receive up to two NDI sources from other devices (increased to 12 in later releases and up to 44 in NewTek's IP Series ).[18]

NDI devices from other vendors followed during 2016. The first third-party products came from Gallery Sienna and included an NDI signal generator for macOS, a desktop scan converter for macOS and the NDICam camera app for iPhone which delivers a native NDI stream from iOS devices.[19][20]

NewTek had previously created a predecessor of NDI called AirSend to get video from external devices into their TriCaster products. AirSend had been implemented by a number of character generator (CG) vendors including VizRT and Chyron. In order to quickly bring these products into the NDI space, NewTek created a new driver to replace the existing AirSend driver, which could be installed on these existing AirSend compatible devices, instantly converting them to NDI compatible devices with no change required by the original CG vendors.[21]

Another early adopter of NDI was VMix, a Windows-based vision mixer which offers NDI inputs and outputs.[22] A significant increase in the NDI installed base came when live streaming application XSplit added support for NDI.[23]

Later in 2016, NewTek delivered NDI 2.0 which added features including support for service discovery across subnets. In April, Magewell announced seamless integration of their PCIe and USB capture devices, allowing access to any video source on the network. This solution created one of the most efficient broadcast video production scenarios possible with unlimited source choice and ubiquitous operating system compatibility.

On 12 July 2017 NewTek announced NDI 3.0 which added multicast, NDI|HX and other new features, introducing support for specific PTZ Cameras with H.264 chipsets and updated firmware.[24]

In April 2018 at the NAB Show, Microsoft announced they had added NDI support to Skype for Content Creators.[25] Other announcements at NAB2018 included NDI 3.5, and new NDI support from vendors including EVS[26] and Avid.[27]

Version 3.4 of FFmpeg added input and output support for NDI when optionally compiled by the end-user to a non-open-source NewTek library.[28] However, NewTek was later found to be redistributing pre-compiled binaries of FFmpeg that incorporated their non-opensource library in violation of the GNU General Public License, causing the FFmpeg project to remove NDI support from their codebase in March 2019.[29][30]

In April 2019, ahead of NAB, NewTek announced the addition of Multi-TCP mode to NDI 4.0 which is reported to leverage hardware TCP acceleration present in silicon, assisting lower-spec processors with NDI transport.[31] NDI 4.0 shipped in September 2019 to coincide with the IBC exhibition.

NDI was designed to work on good quality gigabit local area networks using TCP and Bonjour (mDNS) technologies. In order to work across subnets that don't pass mDNS, NDI supports a mechanism known as NDI Access which allows manual entry of the IP address of machines on other subnets which may be running NDI sources.

Some NDI adopters have run the protocol across fibre connections up to 15 km, although NDI's use of the TCP protocol makes it less suitable for long-distance, high-latency connections due to factors such as bandwidth-delay product and TCP packet loss recovery.[33] The NDI 3.5 libraries now default to a UDP transfer method with forward error correction. This removes the latency requirements that are inherent with TCP connections. NDI|HX uses a lower data rate making it easier to use in bandwidth-limited connections.[24]

To provide an extension of NDI to wide area networks, tools have been developed including the Sienna Cloud for NDI[34] protocol which uses node gateways in each LAN to bridge NDI sources across continents.[35] In July 2017, United States Specialty Sports Association (USSSA) successfully made the first truly wide area network NDI sports coverage using NDI with 5 cameras backhauled across the public internet from a stadium 200 miles away into an NDI based production gallery.[36]

NDI is also used within cloud-based production systems,[37] with its compressed video and unicast transmissions making suitable for cloud-based services like AWS and Azure. in 2019 Sienna demonstrated a complete cloud-based production workflow using NDI, including remote contribution and final publishing via content delivery networks like YouTube Live.

NDI, when running on x86 architectures, requires CPUs that include the SSSE3 instruction set. Most Intel CPU designs starting in 2006 have this instruction set, AMD added support starting in 2011. While not a requirement, NDI will take advantage of Advanced Vector Extensions (AVX) and AVX2 instructions sets for additional performance.

NDI can run on 32-bit or 64-bit CPU architectures, although performance is increased when using 64-bit.

NDI is available with encode support for ARM-based systems, but generally not video decode. Exceptions to this are Sienna NDI Monitor[38] for iOS which does provide NDI decoding on ARM, and Sienna NDI Monitor TV for AppleTV 4k.[39]

NDI|HX devices are typically transmit-only and based on proprietary platforms with hardware H.264 encoder chips. Examples of NDI|HX devices are PTZ cameras, and the NDI Connect Spark SDI to NDI|HX converter box. The NDI stream itself is formed at the stream consumption end via the platform-specific NDI|HX driver, presenting the same uncompressed data to the application software consuming the NDI source regardless of original HX source hardware and compression. NDI|HX2 is a purer implementation of H.264 in NDI packets and will allow software and hardware systems to embed H.264 essence into NDI streams using the NDI Embedded SDK.

NewTek, Birddog and Magewell have all developed FPGA encoding implementations of NDI. In April 2018 NewTek announced the Connect Spark Pro which will be their first FPGA based NDI product. Birddog and Magewell also offer FPGA-based decode of full NDI streams (not NDI|HX). Newtek offers a licensable FPGA reference design for third-party NDI encoders as part of their NDI Embedded SDK.

With NDI 4.0 NewTek announced the addition of a Multi-TCP transport mode. This takes advantage of the hardware TCP acceleration in silicon which helps lower-spec processors handle heavy network load, in contrast to UDP which does not benefit from the hardware acceleration.[40]

NDI supports arbitrary metadata as XML blocks, embedded in video and audio frames, as well as stand-alone metadata frames. The content of these metadata blocks form 3 families.

1. Internal metadata used by NDI invisibly. These messages are used for connectivity and some other fundamental tasks like Tally. They are typically invisible to NDI clients.
2. Defined public metadata. These messages include things like the NDI PTZ Protocol. They are defined by NewTek as part of the NDI SDK.
3. Third-party metadata schemes. These messages are implemented in the same way as defined NDI metadata but the content is based on third-party designs.

There are 8 published examples of third-party NDI metadata schemes:

• In August 2018, Gallery Sienna published a proposal[41] for the carriage of C708 Closed Caption metadata via NDI. It involves the capture of the appropriate line of SDI VANC data, and text-based encoding to be compatible with XML. This is carried as video-frame-based attached metadata. It can be passed to an equivalently compatible decoder for reconversion to SDI along with intact C708 captions. Gallery Sienna implemented this standard into their Sienna NDI SourceMulti product,[42] and the proposal is also being adopted by other NDI product developers. A companion proposal for a universal caption format[43] for NDI has since been published, providing a human readable, rather than binary container for closed captions in NDI Metadata streams.

• In February 2019, Gallery Sienna published a proposal for the carriage of MIDI messages via NDI.[44] This uses real-time stand-alone NDI Metadata frames to carry any sort of MIDI message including MIDI Machine Control, and controller data. NDI MIDI is used within Sienna's NDI processing engine[45] as a generic control protocol between modules and the outside world.

• In April 2019, Gallery Sienna published a proposal for the use of the industry standard iXML Metadata protocol within NDI.[46] iXML allows for unambiguous description of stream contents, and linkage between streams.

• In April 2019, Gallery Sienna published a proposal for the carriage of PTPv2 timing and GMID within NDI.[47] Sienna also implemented this as a Precision Time Protocol converter which captures this data and feeds it as NDI into multiple NDI systems wanting to share this information, such as software to interact with AES67 devices.

• In September 2019, Gallery Sienna published a proposal for the definition and carriage of sub-region stream definition within NDI.[48] This can be used for definition of video wall subcuts of a stream, or for applications like AI analysis of video where a system is following a face in a stream, or following a football - with a pan scan subcut of the stream. This is also used to deliver sub-stream definitions for social media publishing with non-standard aspect ratios - for example, portrait or square video for phones.

• In November 2019, Gallery Sienna published a proposal for the definition of the HDR Metadata description profile for HDR NDI Streams.[49]

• In December 2019, Gallery Sienna published a proposal for the communication of GPI State via NDI connections.[50] This mechanism is implemented in Sienna's GPI Interface module in the NDI Processing Engine[51] which interfaces with the JLCooper Ebox[52] GPI interface, supporting 24 GPI Inputs and Outputs.

• In December 2019, Gallery Sienna published a proposal for the communication of DMX data via NDI connections.[53] This mechanism is implemented in Sienna's ArtNet-DMX Interface module in the NDI Processing Engine[54] which interfaces with the LTECH Artnet-DMX-1 DMX interface, supporting 512 DMX Channels in or out.