By Maria Ingold | iSize Technologies | Published 21st April 2020
I grew up off-grid in a cabin in the New Mexico mountains. That was isolation. By contrast, isolation in the time of coronavirus is incredibly connected. While working, socialising and relaxing from home have impacted that connectivity, new patterns are emerging as well as opportunities for the future.
What is the scope of the impact?
Akamai, a content delivery network (CDN), saw global internet traffic increase by 30% in March, an entire year’s growth, and without live sports streaming. Comcast saw a 32% increase in peak USA traffic over March, with plateaus in early lockdown markets.
Even before COVID-19 video was 60% of downstream internet traffic. When Conviva analysed three weeks in mid-March they discovered that video streaming viewing hours jumped more than 20% globally in that last week, up 27% in the USA. By the end of March, Comcast saw a 38% USA increase.
Although Internet service providers (ISPs) and CDNs are engineered to deal with peak changes, when usage spiked, the European Commissioner asked streamers to switch to Standard Definition (SD) when High Definition (HD) wasn’t necessary. The European Broadcasting Union (EBU) then issued recommendations for adapting streaming quality during times of crisis.
Assuaging one concern, Conviva discovered that daytime viewing jumped nearly 40%, spreading peak load, but that still leaves lots of bits flowing across the internet.
Netflix and Google’s YouTube agreed to reduce bitrates in Europe for 30 days, with Netflix dropping by 25% and YouTube moving to SD as a default globally. Both were crucial, because while Netflix usually has the largest percentage of video traffic, YouTube is currently generating almost twice the traffic of Netflix. Amazon Prime Video, Apple TV+ and Walt Disney’s Disney+ soon followed.
Consumers were concerned. They were paying for HD but would get SD. Netflix explained that customers would still get the SD, HD and Ultra-High Definition (UHD) resolutions they paid for, just no longer the highest quality from a “bitrate ladder” of low to high bitrates and resolutions.
What are the long-term opportunities?
Netflix’s total energy consumption for 2019–451,000 megawatt-hours — could power 40,000 average American homes for a year, at an 84% increase over 2018, compared to 20% user growth.
Netflix has 167m subscribers. Disney+ has 50m, with 226m subs predicted by 2024. Reducing bits creates a more sustainable energy-consumption to user-growth ratio and helps companies meet their environmental impact objectives.
During the 30-days of COVID-19-inspired bitrate reduction, streamers will have saved money by reducing storage, distribution and energy costs. If one million people watch one hour per day, at 1 GB of data per hour (somewhere between SD and 720p HD), and it costs .0025 USD to stream 1 GB to one person, that’s nearly $1 million per year ($912,500). YouTubers watch one billion hours per day. That’s nearly $1 billion per year. A 25% savings is $228 million.
While these short-term actions enabled quick bitrate reductions and increased margins, they don’t preserve quality. Consumers won’t tolerate that indefinitely.
How to cut costs and maintain customer satisfaction
A codec encodes (usually in hardware) the moving image source and decodes (usually in software) on a device to display it. It reduces bitrate as much as possible while attempting to maintain fidelity. Codecs range from older, widely supported but higher bitrate like MPEG-4 AVC (H.264) to newer, less supported, time and power-hungry but lower bitrate ones.
Per-title encoding is a bitrate reduction tactic pioneered in 2015 by Netflix. To measure fidelity, Netflix used quality metric PSNR (Peak Signal-To-Noise Ratio), but it doesn’t always measure how it looks to a person. Neither does SSIM (structural similarity), designed to improve on PSNR. So, Netflix co-created VMAF (Video Multi-Method Assessment Fusion), a perceptual quality metric.
Machine learning (ML) can reverse engineer perceptual metrics to make encoding more effective. When this precedes encoding — precoding — it works with any codec, encoder and decoder. There are trade-offs between the sharpness of VMAF, which can look artificial, the naturalness of PSNR and SSIM, and the blurriness and lack of fidelity caused by reducing bitrate.
I advise iSIZE, a machine learning precoder that claims 20%-40% bitrate savings (up to 60%) without changing the resolution and typically improving VMAF. Latency is one frame. I asked expert reviewer Jan Ozer to independently test iSIZE’s BitSave product. He tested using the MPEG AVC (H.264) codec.
Jan confirmed that “BitSave is a legitimate processing technology and not a [VMAF] hack”. Ultimately, “[a]fter many hours of testing, [Jan] found that BitSave’s technology is valid and valuable” though he recommends subjective testing. I agree and recommend testing various bitrate savings and metric balances. Regardless of the solution you choose, remember to balance long-term sustainability and cost-cutting with perceived customer experience.