It is 2021, let us just say it, video streaming and in particular video transcoding is a big deal. In order to meet that demand, there are a lot of different infrastructure options available. CPUs and GPUs have been able to do video transcoding for some time. Google has its Google YouTube VCU for Warehouse-scale Video Acceleration but that is a customized hardware and software stack. Amazon AWS EC2 instances bring Xilinx U30 FPGAs to the cloud and lower the cost of transcoding versus CPU and GPU options.
Amazon AWS EC2 VT1 Instances
First off, let us get to what a Xilinx U30 is since those are really the stars of the VT1 offering. The Xilinx U30 is the Alveo U30. We covered this previously in our Xilinx Live Video Transcoding Product Line Launch article. The Xilinx Alveo U30 is designed to be the lower power and higher density solution with the Alveo U50 being the higher-end solution.
Here are the key specs for the card. Key here is that it is a 40W PCIe card so the idea is that a company, like Amazon AWS, can put many of these cards into a server and achieve higher density. The higher density and lower power mean that one can get faster transcoding at a lower cost to AWS and to the user. AWS says that this solution is 30-60% lower cost than its CPU or GPU transcoding offerings. It is also something that in the future would be the target of being an EDSFF form factor accelerator at 40W, although there are no announcements of that type of product at this time.
Xilinx says it can transcode 4K video at 60fps across two streams in real-time. It can also support lower-quality transcodes. For further bandwidth savings, the U30 can also transcode HEVC. One can see some additional features on the U30 half of the slide below.
As part of this offering, Xilinx not only offers the FPGA-on-a-PCIe card hardware solution but also a software solution. Xilinx has the ability to add a single line into ffmpeg, which is perhaps the default transcoding application these days, and enable the Alveo U30 to do the transcoding work. Something that Xilinx got right here is making a change like this relatively trivial to implement and therefore relatively easy to integrate into a workflow. Amazon can ensure the FPGA is initialized and available to the instance as part of its service offering. Xilinx has the software to make it relatively easy to integrate into a widely used industry tool.
In terms of actual instance sizes, the AWS EC2 VT1 instances come in three sizes with perhaps the most important feature being the number of cards, in 1, 2, or 8 accelerator configurations.
|Instance Size||Xilinx U30’s||vCPU||Memory (GiB)||Network Bandwidth (Gbps)||EBS Bandwidth
|1080p60 Streams||4Kp60 Streams|
|vt1.3xlarge||1||12||24||3.125||Up to 4.75||8||2|
In terms of pricing here is the associated breakdown at launch:
|Instance Size||On-demand (Price/hr)||1-yr Reserved Instance Effective Hourly||3-yr Reserved Instance Effective Hourly|
As one can see, the accelerator to stream recommendations aligns with the U30 that we showed last year on STH, and that are in the slides from that announcement. Also, for those that need a constant transcoding capacity, the reserved instance pricing saves a lot here.
Our general sense is that we will see more of these types of solutions in the future. Now that Google has started discussing its transcoding solution publicly, it will raise industry awareness around the benefits of offloading transcoding to smaller-scale organizations. For those using AWS, and that do not have the scale to build their own ASICs for warehouse-scale transcoding, the new EC2 VT1 instances offer something cool.
AWS has something else that may make folks excited. The company also announced that it will be bringing these Alveo U30 VT1 instance types to AWS Outposts for on-prem deployments. Think of using an outpost for live event streaming via an AWS Outpost as an example. It seems like AWS has bigger plans for the U30 than just its cloud data centers.