Distribution and streaming – With rare exceptions , like 70 mm IMAX , movies and television shows are delivered digitally to cinemas , TVs , and phones all over the world . AI plays a major role in video compression and data optimisation from the cloud to packet processing and bit rates as video streams across wired , Wi- Fi , and 4G / 5G networks .
Captioning , translation , and localization – GenAI services can combine automated speech-to-text with response generation to produce closed captioning , transcripts , and translations on the fly . AI also assists in reformatting to match local broadcast specifications , frame rates , and device aspect ratios .
Marketing – AI helps identify potential hits , trending songs , and hot shows , then places them in front of the right audience , at the right time , on the right device . Thanks to AI , streaming platforms ’ recommendation engines combine a nuanced understanding of audience behaviours and preferences with real-time analysis of what ’ s hot .
General business intelligence – AI and GenAI services create value in both directions . The users of AI services receive help , have more efficient workflows , and produce labour-saving work . The service provider receives intelligence . Wherever AI supports a function or role , the business receives data about that role that GenAI can turn into knowledge and action .
The same GenAI service that helps a line producer optimise shooting schedules can review fleet data and map more efficient routes for studio trucks . It can scout locations to minimise travel , track carbon emissions , and calculate offsets all while optimising cloud computing resources to the enterprise ’ s real-time needs .
As GenAI spreads in media and entertainment , will the net effect be good or bad for sustainability ? At this stage , the trade-offs are difficult to calculate . We only have estimates of GenAI energy use and a limited view into how rapidly studios , streamers , and infrastructure providers are adopting GenAI . However , we can make some educated guesses .
Replacing location shoots with virtual , AIgenerated sets should reduce net energy use . Traveling to a location means moving talent , crew , and equipment . Powering a set usually means diesel generators , although there are greener alternatives that run on propane or natural gas . Even with green shooting practices in place , a location shoot is an energy and timeconsuming proposition .
Using AI to improve file compression , encoding , and transmission should make file transfers and streaming faster , less expensive , and less energy-intensive . Compared to printing and shipping thousands of reels of film or Blurays , digital delivery is clearly more sustainable . GenAI promises to make it even more efficient .
General operational efficiency — from faster post-production to insight into a project ’ s carbon footprint — should improve steadily as GenAI takes on more roles and functions . Over time , automated , continuous improvements in processes should add up to significant energy savings .
Whether these hypothetical efficiency gains outweigh the energy cost of GenAI is an open question . One way to make sure GenAI does have a positive impact is to improve the energy efficiency of model training and deployment .
How to make LLMs and GenAI more sustainable The energy costs of GenAI are built-in to the cost of the service , which makes GenAI ’ s energy impact opaque to the end user . Changing the energy equation for AI rests with the companies providing AI services and manufacturers like HPE who build the supercomputers and data centres that power GenAI . There are levers we can pull to reduce the energy overhead of GenAI , and HPE is working to create new ways to make GenAI sustainable for every industry .
Model training takes massive amounts of time , computing power , and electricity . Every improvement we make in the training process pays immense dividends . HPE has developed the HPE Machine Learning Development Environment . It includes tools for tuning training workloads , so they use hardware more efficiently . The environment runs on machinelearning-specific accelerators that are up to 5x more efficient than off-the-shelf systems . 2
Even though model training is energy-intensive , AI services consume 90 % of the energy used to deliver AI . 2 Where those services run can be a major factor in energy consumption and carbon footprint . For example , the average private data centre can be half as efficient as a cloud data center . 2 Newer , more energy-efficient servers like the HPE ProLiant Gen11 come equipped with workload-specific accelerators that can deliver up to 10x better performance per watt . 4
The physical location of a data center matters , too . It takes far more energy to move electricity than it takes to move data . Locating data centers near power generation facilities minimizes energy lost to power transmission . If those facilities use solar , wind , or hydropower , the data center ’ s carbon footprint will be significantly reduced . This is why many hyperscalers and service providers are shifting data center infrastructure to regions with abundant hydropower and colder climates that can reuse the data center ’ s waste heat .
Optimizing AI models to run more efficiently can also drastically improve energy performance , regardless of what hardware these models run on . Sparsely activated deep neural networks can consume < 1 / 10th the energy of similarly dense deep neural networks ( DNNs ) without sacrificing accuracy . 2
Taken together , increasing supply-side efficiencies through optimization , more efficient hardware , and renewable energy can collectively reduce AI carbon emissions 1000x . 2 With efficiency gains like that , GenAI in media and entertainment — and other industries — may prove to be a gain for sustainability . At HPE , we are working to ensure that this is the case .
Find out more about how we can help you build solutions that are both sustainable and AIdriven at hpe . com / solutions / oem .
1 https :// www . tomshardware . com / tech-industry / nvidias-h100-gpus-willconsume-more-power-than-somecountries-each-gpu-consumes-700w-ofpower-35-million-are-expected-to-besold-in-the-coming-year
2
David Patterson , et . al , Carbon Emissions and Large Neural Network Training , arxiv . org , accessed April 2024 .
3
Sarah McQuate , Q & A : UW researcher discusses just how much energy ChatGPT uses , University of Washington News , July 27 , 2023 , accessed April 2024 .
4
~ 10X better performance / watt over the NVIDIA L4 submission for the Computer Vision models ( ResNet and RetinaNet ).
EUROMEDIA 15