Video transcoding is a huge space. We covered the Google VCU Video Coding Unit A few times. Xilinx, now AMD has been popular with services like Twitch in the past. With so much Internet traffic being video, there is a huge opportunity to improve codecs and lower bandwidth demands. Further, the number of real-time streams is increasing with streaming from PCs and mobile devices becoming mainstream. The AMD Alveo MA35D is AMD’s new custom silicon solution for video transcoding.
AMD Alveo MA35D Custom Silicon for Video Transcoding
The AMD Alveo MA35D is part of the company’s media accelerator line “MA”. This is the start of what will be a line for AMD, not just a single-card solution. Today’s card has two chips on a single low profile card and a $1595 MSRP.
We asked about that connector on the back edge of the card and were told it is a debug connector that may not make it to the final version, but it is in the renderings.
On a generational basis, this new MA35D is the update to the AMD Xilinx Alveo U30. This accelerator uses hardened logic and achieves significantly more performance.
Here is the 5nm ASIC. The Alveo MA35D supports AV1, VP9, h265, and h264 and up to 8K resolution. Something that is different with the ASIC is that the VQ QoE Engine and VQ look-ahead features can, at low latency, make determinations of quality/ compression impacts and optimize streams in almost realtime.
On the codec side, AMD is not only supporting AV1, but is new codec improvements help increase compression thus lowering bandwidth of streams.
Since the Alveo line comes from the Xilinx acquisition, the Alveo MA35D has AI inference blocks as well. AMD showed an example of AI inference recognizing faces as important parts of streams and therefore prioritizing quality in that part of the frame.
Here is AMD’s slide showing the pipeline to help reduce bitrate and bandwidth with the new cards that can handle 1080p60 streams at 1W.
One of the more interesting slides is this one. AMD’s approach is to enable not just CPU transcode for higher-end streams (e.g. the Super Bowl), but also lower $/stream applications. While the hardware cost and power cost may not be large in the context of the Super Bowl, it is in terms of the Twitch streamer that is just starting out.
Aside from lowering costs per stream, the other goal of the MA35D is to help layer value-added services for ads and micro-transactions on streams with its partners.
We asked AMD and this is not the type of card meant to be purchased one or two at a time. Instead, this is meant to be sold to AMD’s large streaming partners. Much of today’s transcoding happens when a GPU is used just for their video codecs. AMD is looking to make this much more efficient.