Production & Post Global Viewpoint – July 2020

GPUs have their history in PC gaming. Originally, software developers would write image processing software that ran directly on the CPU, but as the quality and interactivity of the games improved, the load on the CPU became intolerable. One solution to this was the development of GPUs to provide hardware accelerated parallel processing that took away some of the intensive image processing from the CPU.

Advances in machine learning (ML), a branch of AI, have had some interesting and possibly unintended consequences, that is, the rapid development of GPUs. ML uses similar mathematical techniques to those used in image processing. Linear algebra is used extensively in science and image processing and machine learning are no different. The parallel processing offered by GPUs, specifically for matrix and vector functions and the highly recursive back propagation used by recurrent neural networks, significantly speeds up ML learning.

It’s fair to say that ML is a massive area of interest for academics, engineers and those who can see the positive impact ML will have on all our lives. Much of this research is outside the broadcasting domain but it does have significant applications in broadcasting. Image recognition and metadata tagging for library archiving is just one example.

Consequently, there is a massive demand for GPU cards as the hardware accelerated parallel processing functions not only lend themselves to image processing, but also to ML learning. This has paved the way for vendors to invest heavily in GPU development and provide some truly eye watering processing resource and reliability.

Cloud service providers and COTS hardware vendors have known this for a long time and it’s no coincidence that we are seeing many of them providing GPU solutions.

One question I constantly ask is what does it mean to be COTS? There are some formal definitions that are derived from government procurement specifications, but as a guide, what is a COTS service or product?

To my mind, COTS is the opposite of bespoke or custom. As the name suggests it can be purchased “off the shelf”. However, as many broadcasters have found, the “shelf” can be very high in terms of cost. This doesn’t mean the COTS product or service cannot be highly tuned or adapted to a specific application, just that it is not a one-off, or very small batch run. Microsoft is a COTS provider and they are only one supplier, but the COTS products they provide are developed and supported by thousands of people resulting in the much-reduced possibility of a single point of failure, both in the product functionality and supply chain.

Based on this finger-in-the-air definition, I believe GPUs have now entered the world of COTS. GPU and server vendors are actively collaborating to provide off-the-shelf products and a quick glance at the many cloud providers will see many GPU options. The single point of failure for procurement and product functionality is greatly reduced.

This provides a fantastic opportunity for broadcast vendors. GPUs have many applications in ML as well as image processing and our demands for real-time low-latency video, audio and metadata distribution and processing can be more easily satisfied. Once again broadcasters are benefiting from developments in other industries and I’m looking forward to seeing a whole host of COTS GPU applications taking the broadcast world by storm.