AMD builds breakthrough AV1 encoder chip for streaming services

When AMD bought Xilinx, it acquired a slew of new technologies and entrance into markets developed by the latter company. One of those unique technology niches includes high-performance video encoders for streaming media companies. These types of companies have massive bandwidth and data storage requirements. To reduce these requirements, companies compress video, increasingly with an emerging new format called AV1 (see sidebar below). AMD’s latest video chip is designed to stream AV1 in the in a power efficient and economical way.

In 2019, prior to AMD’s acquisition, Xilinx bought a company called NGCodec, which specialized in low-latency cloud video encoding using Xilinx FPGA acceleration to provide high-quality live streams at low bit rates. The NGCodec IP supported H.265/HEVC and VP9 standards and had promised future support for H.264/AVC and AV1. One of the NGCodec customers was Amazon’s Twitch game and event streaming service, which used VP9 encoding.

This business has turned out to be a significant opportunity, and to follow up on the earlier success with NGCodec IP and the Alveo U30 media accelerator card designed for the streaming market, AMD’s Adaptive and Embedded Computing Group (AECG) has developed a purpose-built application-specific integrated circuit (ASIC) for online streaming video using a new AV1 codec. This new card is the Alveo MA35D and it delivers a massive leap in throughput and compression over the U30 with the AV1 codec support.

What is an AV1 Codec?

AV1 is a powerful video encoder-decoder (commonly called a codec). Using more advanced video codecs can reduce the bandwidth needed to send video to the cloud and from the cloud to users. The file sizes of uncompressed videos from cameras are very large and increase as the quality and resolution of the videos increase. Without compression, videos require enormous and uneconomical amounts of storage.

The video codec is a specialized algorithm that examines each frame of a video and works to reduce its size by eliminating features that the human eye can’t perceive. They can also further reduce bit rate requirements by producing additional intermediate frames, using information from reference frames and motion vectors of key features. By using these algorithms, the size of the videos comes down without reducing the perceived quality. These codecs are essential to stream high-quality video over the internet.

Different compression algorithms are used for different types of video streaming, and each one offers different efficiencies when reducing video size. Each also have different processing requirements.

The AV1 codec was developed by the non-profit Alliance for Open Media, and doesn’t require a license. Due to this open nature, the codec can be used by anyone without paying royalties to the developers, unlike the H.264 and H.265 codecs. Founded by Intel, Amazon, Apple, Netflix, and others, the Alliance of Open Media released AV1 in 2018. The AV1 codec is supported by many recent SoCs and GPUs. When it comes to compressibility/image quality evaluations, AV1 leads the pack today.

The performance of the MA35D is more important than ever- as the video market evolves from massive broadcast streaming content from a central server (a one-to-many model) to a distributed model with many video streams coming into and out of a multi-tier cloud architecture (many-to-many model) driven by online meetings, video conferences, massively multiplayer online games, social media streams, and soon augmented reality (AR) headsets.

This new massively distributed content is driving demand for bandwidth, leading to increased costs while choking network bandwidth and clogging cloud architectures.

The MA35D has up to 4x the channel density compared to the Alveo U30 card. With support for AV1, the MA35D can offer 1,.8x higher compression over H.264 compression standard running on the U30. The MA35D running AV1 can also deliver 4x lower latency than the U30 running H.264 at 4K resolutions.

At the heart of the chip are two custom AV1 encoder blocks. The company claims these AV1 encoders are the industry’s most powerful codecs and deliver lower bit rates. Because the chip is built using a 5nm process technology node and is designed using fully custom logic, the chip is extremely efficient. It can transcode up to 32 1080P/60 bitstreams at only 1W per channel. There are two custom chips per Alveo MA35D card. Each chip has four encoders – two high performance AV1 encoders and two multi standard encoders that support H.264, VP9, and AV1.

The chip also adds an AI block (because, of course, they had to add AI) for AI-powered quality enhancements. The AI feature can be used to identify critical parts of a scene to apply extra image precision while applying lower bitrate techniques to less critical areas. Such optimization techniques include region-of-interest (ROI) encoding for text and face resolution to provide clearer images and video artifact detection to correct scenes with fast motion and high complexity.

Inside the video pipeline are hardened units for decoding an incoming video stream, scaling and compositing images, rescaling video bit rates, predictive image optimization, and finally encoding the video output using the appropriate compression standard. In the U30, these functions were configured out of FPGA blocks. Implementing these features in a custom SoC increases performance while reducing power consumption.

Other features of the chip includes: dual set of quad-core RISC-V processors, and an LPDDR5 memory controller, and a PCIe 5.0 interface to the host processor

The new MA35D card is designed to be software compatible with the open-source U30 card’s software. Video frameworks supported include FFmpeg, Gstreamer, and C/C++ APIs from AMD. The Alveo MA35D card is sampling now and will be available in production in Q3 of this year for a list price of $1595. The Alveo MA35D will be displayed in the AMD booth at NAB, North Hall N2158. Forbes