Chipmaker AMD debuted a new media accelerator card that promises to help enable interactive live streaming at scale, while providing capex and opex savings for service providers in the face of changing infrastructure and economic requirements needed to support emerging applications.
The AMD Alveo MA35D is the industry’s first 5nm ASIC-based media accelerator, according to AMD, powered by a purpose-built video processing units (VPUs) that support the AV1 compression standard.
The new product comes with a bevy of features such as increased channel density support – needed for two-way interactive streaming such as with cloud gaming, social streaming, and live shopping, among others – with 32x 1080p60 streams per card, improved power efficiency and ultra-low latency performance.
The Alveo MA35D is targeting customers at the cloud service level, but Sean Gardner, head of Video Strategy and Market Development AMD’s Video Transcoding Group, noted the significance as a fundamental building block that impacts, at the service provider level, how much it will cost those operators to deliver new applications and use cases.
Part of what’s driving the need for the product is the proliferation of streaming and particularly an increase in livestreams, AMD’s Vincent Fung, senior product marketing manager, told Fierce. The company cited data showing that 70% of the global video market was comprised of live video in 2021 (with the remaining allocated to on-demand content) – with many live streaming platforms such as YouTube, Facebook, Twitch and others gaining popularity during the height of the Covid pandemic.
Interactive live streams break existing infrastructure models
In building its new accelerator card, AMD aims to address shifts in economic requirements tied to changes to the existing infrastructure model that’s needed to support new applications.
Existing infrastructure for the traditional live broadcast model, such as the Super Bowl for example, follows a one-to-millions format – where one stream reaches out to many but delivery is asymmetrical and the provider doesn’t have to worry about a concurrent inflow of traffic.
However, AMD sees its Alveo MA35D powering a variety of new live stream video use cases including live events, collaborative video tools, social streaming of user-generated content, and watch parties. This wave of applications breaks the existing centralized data center model, Gardner said.
Take a scenario such as Twitch, for example, where a large number of viewers are also possibly participating with their own stream from separate locations, so instead of 1-to-millions it becomes many-to-many. (Fung noted that essentially “every participant becomes a broadcaster”) and potentially getting a personalized stream delivered – also leading to the previously mentioned shifts in economic requirements.
“As those scale, the number of videos that need to be processed…and transcoded drastically increase,” Fung said. “So we need to have a solution that actually makes economical sense for these operators to be sustainable.”
Fundamentally, interactivity means low latency, according to Gardner. And once a provider is delivering interactive, that means it’s no longer a multi-cast – where one is broadcasting a single channel and everyone sees the same thing.
“Now you’re treating customized channels on a per group or per user level,” Gardner noted.
With these applications, traffic increases drastically, now with a huge volume of streams coming on the ingest side – with a “massive volume of channels to process” as well as send out, putting strain on infrastructure in terms of servers and the network. As companies seek to offer more interactive live streaming services there becomes a need to push closer to the end user and cut down latency as less lag time for the round trip becomes critical. And these needs mean the centralized datacenter model no longer holds as video processing power moves into edge datacenters, often located nearby to more densely populated areas.
While working at the chip level, the work AMD has done with customers has been focused on their future needs, Gardner said, mapping back to the architecture to ensure that it will support the ability to achieve the right economics to deploy those emerging applications at scale, at the right price point to be able to be profitable.
Maximizing limited footprint
A key challenge is that datacenter space is limited - particularly those that are confined to smaller square footage near large population centers – and AMD sees one of the benefits its Alveo MA35D brings is more bang for the proverbial (and actual) buck or lower cost on a square footage basis.
Gardner noted that as providers get closer to the edge or user to get the needed latency required “there is an inverse impact on cost per square footage” which “is massive,” particularly as datacenters or edge centers get into higher density population areas like Manhattan or downtown LA with more limited footprints space-wise.
One aim with the new accelerator is maximizing that limited space and power budget in a datacenter in a cost-effective way. With a processor that is smaller, more power and energy efficient, Gardner pointed to both capex and opex savings for service providers. The smaller accelerator card is able to take up less rack space in data centers while delivering a greater number of high-quality streams. For example, AMD said with up to 32x streams per card at 1 watt per stream, a 1U rack server equipped with eight cards delivers 256 channels, maximizing the number of streams per server, rack or data center.
“We worked closely with our customers and partners to understand not just their technical requirements, but their infrastructure challenges in deploying high-volume, interactive streaming services profitably,” said Dan Gibbons, general manager of AECG Data Center Group, AMD, in a statement. “We developed the Alveo MA35D with an ASIC architecture tailored to meet the bespoke needs of these providers to reduce both capital and operating expenses for delivering immersive experiences to their users and content creators at scale.”
On the capex side, Gardner said the biggest expense is bandwidth and initial investment when it comes to space and related cost per channel, in terms of hardware. He noted that the new purpose-built video processing unit features next-gen AV1 transcoder engines that help deliver up to 52% bandwidth per channel savings.
While AMD has done TCO (total cost of ownership) analysis with some of its customers, Gardner noted total savings depend on their respective starting points, although he cited significant savings from a compression and power perspective.
“It is very customer and use-case dependent, based on where their starting point is, but we’re talking about 10s of millions of dollars [in capex/opex savings] on an annual basis,” he noted, adding that in some cases it could be hundreds of millions of dollars over the life of a program coupled with the ability to scale.
When it comes to live video, latency isn’t the only critical factor as video quality and the visual experience is also paramount. Fung noted that AMD is also bringing an AI-enabled video pipeline to address video quality in a dynamic way.
Compared to previous generations, AMD's new accelerator boasts 1.8x greater compression efficiency, 4x lower latency, and 4x higher channel density. The company will be showcasing the MA35D at NAB in Las Vegas this month. It’s currently sampling now with early lead partners and customers, with production shipments expected in Q3.
An earlier version of this article said AMD's new accelerator boasts 2x lower power cost per channel, and has been updated to instead reflect a 1.8x improvement in compression efficiency.