The artist formerly known as Xilinx has unveiled its first new product since its acquisition by AMD, with the Video Transcoding Group launching the Alveo MA35D Media Accelerator. The plug-in card is apparently an industry-first 5nm architecture, and is being aimed at the new crop of interactive and dynamic video streaming workloads.
Notably, this is an ASIC-based approach, and not the FPGAs that Xilinx was best known for. The reprogrammability of the FPGA was often its main selling point, with the ASIC often criticized in that sales pitch for its relative lack of flexibility. The trade-off was that the FPGA was more expensive, but if you knew that the ASIC’s workload would not change during its lifetime, then the cheaper specialized ASIC approach could be used.
Initially, the ASIC implementation could sound anathema to the interactive streaming focus. However, at the end of the day, video is video, and so transcoding does not really care what content it is handling. Accordingly, the Alveo MA35D, thanks to its ASIC design, lacks support for VVC (H.266) – focusing on the more set-in-stone AVC, HEVC, and AV1.
However, while the video is not subject to much change, AMD expects the deployment architectures of the video workflows to shift significantly – to support the expected move to interactive media services. Examples of these include live events, cloud gaming, live shopping, watch parties, betting, e-learning, social streaming, remote production, collaborative video sessions, live auctions, and customer service.
As such, the sales pitch for the MA35D points to the current streaming model’s one-to-millions construction, where the live feeds from a sporting event pass through contribution and production workflows, before reaching the origin server, and then passing through CDNs and edge servers to reach the end-viewer.
With the aforementioned examples, AMD is banking on there being many more ingest and egress streams, which will significantly increase the channel density of an ISP network.
This is where the card-based approach comes into its own. The claim here is that one server housing eight MA35D cards is equivalent to many more servers relying on CPU processing power alone.
This is for a 1080p60 stream, using x264 (AVC). AMD claims this represents a significant capex saving, as well as plentiful opex savings in terms of power per channel, and 52% less bandwidth per channel. The MSRP for the MA35D is $1,595, with a 50 watt TDP (Thermal Design Power – how much electricity you can safely pump into it, based on its passive cooling capabilities).
Each MA35D can transcode up to 32 channels of 1080p60, using just 35 watts of power – the source of the one watt per channel claim. Each MA35D houses two of the Video Processing Unit (VPU) ASICs. A maximum latency for a 4Kp60 encode of 8ms is also advertised, as well as support for 8Kp30.
With eight cards in a server, that provides 256 channels at 280 watts, and as you can fit many servers in a cabinet, this should add up to significant space and power savings in both data centers and edgier deployments.
It is a fairly damning indictment of the HEVC patent pool process that the vast majority of video delivered today is still using AVC. To this end, while the AV1 comparison that AMD makes in the announcement initially looks unfair, it is a sad reality that HEVC is not the main draw.
Here, AMD claims best-in-class compression, of up to 52% for the MA35D when using AV1, compared to the baseline X264 Very Fast encoder. Using AVC, the MA35D provides a 24% improvement on the baseline, which is likely enough to tempt those stuck on AVC, but it is notable that AV1 is still not lumped in with the “mainstream codecs” crowd in the description. With performance uplifts like this, however, this positioning should hopefully begin to change.
A pre-release briefing Video Multimethod Assessment Fusion (VMAF) comparison suggests impressive gains on HEVC and AVC, which compound to create a substantial difference at scale. However, the numbers in the claims shifted, and AMD dropped the precise figures from its announcement.
There is a new feature that uses an AI function to prioritize transcoding efforts to the areas that the viewer is most likely to notice. The example given is targeting faces, which does appear to significantly boost the quality of the faces at a claimed “comparable” bandwidth. Of course, as with all such claims, one must be wary of cherry-picked results and best-case scenarios.
As for the changing market, AMD is still positioning its Epyc CPU line as the best choice for streaming workloads that have the least ingest or source streams but the largest audiences. The MA35D and its ilk are positioned once the number of source streams seems to pass 10,000 – comprising live sports betting, interactive media, cloud gaming, and live eCommerce.
Here, the scale appears to stretch from 10,000 to 10 million, which appears like a staggering volume today. However, the video industry has been well and truly fractured in this regard. Gone are the days of three broadcast channels and a captive audience. To this end, the largest audiences are expected to be monetized by advertising, while the smaller audiences will rely on micro-transactions, according to AMD’s projections.
At the time of writing, Wowza had just announced that it added the Alveo U30 Media Accelerator card into its Wowza Streaming Engine software offering. Demonstrations will be at NAB, and Wowza claims this will boost the environmental performance of its services. However, that announcement was lacking numbers on this front.
While announced this week, the Alveo MA35D Media Accelerator will not enter full production until later this year, probably Q3.