Services Global Viewpoint – December 2021
Empowering Floating Point With GPUs
GPU software processing is not only improving time to market for vendors but is also advancing the quality of video and audio processing.
SDI and AES has served us well, and no doubt will continue to do so for years to come. However, the integer sample representation of SDI has led us down a path of truncated data types and with it, compromised quality. Key to improved quality is the ability to process floating point samples as this provides a much deeper sample resolution to help reduce aliasing and other artefacts.
It's fair to say that FPGA vendors have fought back with several manufacturers providing onboard floating-point processing engines directly on the silicon. But the development times and support costs of custom hardware are arduous to say the least. Even if a vendor builds their own generic FPGA board that is “future proofed” it’s invariably very difficult to predict trends in such a high paced sector as the broadcast industry.
I have spent many years involved in the development of FPGA systems and I know from experience that the utopian future proof board is somewhat illusive. Either we build an FPGA board that is so future proof that it’s not commercially viable today, or we build an FPGA board that will provide all our needs now, is commercially viable, but will be out of date in twelve months’ time. Thus, requiring a new FPGA hardware development cycle before the last one has even finished.
GPU resources continue to deliver much increased capacity, data throughput and processing speed, so they can now lend themselves to broadcast applications, with some to spare. And the flexibility with which they can be programmed and supported is compelling. Off-the-shelf cards are available within a few days of placing an order and we can be assured that not only do GPU vendors have massive buying power to deliver their products, but they are designing a whole new generation of GPU cards that are further improving on the current generation with incredible backwards compatibility.
The experience being gained in other industries is only helping the development of the GPU, especially with medical, where attention to color accuracy and real-time delivery is a must, and finance where nano-trading is driving IP delivery and processing to new levels. Another unseen consequence of GPU based solutions is that there is no reason why broadcast vendors cannot adapt their products to other industries, such as medical, where our ability to process hi-res images is second to none.
Floating point processing delivers incredible video quality as we have much greater dynamic range resulting in the noise floor being significantly lower. Also, video compression benefits greatly as the sample truncation is almost un-noticeable, and this will only improve as the processing speed and bit depth with which GPUs process will continue to expand as vendors plough more resource into their development.