Why Does AI Consume So Much Energy?
While there should be little doubt Artificial Intelligence (AI) is here to stay and will continue to rethink how we work and play, perhaps a quick pause is necessary to surface a very important question that threatens to burst the AI bubble: Why does AI consume so much energy?
To get to the bottom of this thorny question, I was able to catch up with Lyline Lim, Head of Impact and Sustainability at PhotoRoom, a leading AI photo editor. According to Lim, AI models consume so much energy because of the vast amount of data that the model is trained on, the complexity of the model, and the volume of requests made to the AI by users.
AI Training and Energy Consumption
During training, the AI model “learns” how to behave based on a large set of examples and data. Training an AI model can take anywhere from a few minutes to several months depending on the amount of data and complexity of the model. During this time, GPUs – a type of electronic chip used to process large amounts of data – are running 24 hours per day, consuming a large amount of energy.
“The more complex the model and bigger the dataset, the more energy the AI will use during training,” said Lim.
Inference and Energy Usage
Another reason for energy consumption is when the AI is making an inference, the process of answering users’ queries. The AI first “understands” the query then “thinks” of an answer before sharing the conclusion to the user. Each AI inference requires GPU processing power, which uses energy. “The more popular the AI model, the more inferences will be run, and the more energy will be consumed,” said Lim.
Making AI functionality more sustainable
For most companies, the biggest incentive to make AI functionality more sustainable is cost and user experience. The AI industry is relatively new, and at the stage where companies are looking for quality rather than cost or speed of execution.
“There is so much growth and investment that cost is still a secondary for consideration for most businesses in the AI space,” said Lim. As more AI models require more and more GPUs, the GPU providers may struggle to keep up. This increase in demand might eventually lead to providers increasing the cost of their GPUs, which would force AI companies to use GPUs more efficiently.
Over the long term, AI companies will also be vulnerable to any increase in energy costs, also incentivizing businesses to become more cost and energy efficient.
Environmental impact of AI tools
In the AI space, environmental impact is highly related to cost-efficiency and user experience which are central topics for any business. Lim believes there is no doubt AI innovators all along the chain will take the topic seriously as the AI market matures, regardless of their initial care for the environment.
“Reducing the environmental impact of AI tools is tightly aligned with improving the user experience, and cost efficiency,” said Lim. “If your model is slower, you will spend more on computing power, and the user will suffer from a slower experience. So, if you can work on making your model faster, your user will have a much better experience, you will reduce your costs and reduce energy consumption. Making AI models cost-effective is aligned with reducing environmental impact and providing the best experience for the user.”
Specialist AI models to the rescue?
According to Lim, generic AI models require much broader, larger datasets to be trained than specialist ones, and therefore consume far greater magnitudes of energy.
“By comparison, specialist AI models like PhotoRoom, which is tailor-made for product photography, consume far less energy,” said Lim. “We did the math and found that PhotoRoom consumes 164 times less energy than a generalist image model like Stable Diffusion XL.” Lim said the focus for PhotoRoom has always been about using technology in ways that are useful and accessible to users.
“We want to help more people make amazing photos–whether that’s a stay-at-home mom creating and selling her own jewelry part time who would never be able to afford a professional photographer, or an ecommerce team who are trying to automate manual tasks and simplify their workflows.”
PhotoRoom started by creating a phone and desktop app that could remove backgrounds better than anything else on the market, but quickly realized how big the opportunity for generative AI was going to be. Last year PhotoRoom built the first version of what became Instant Backgrounds, its generative AI background creator.
“This was a very successful bet for us,” said Lim. “We were the first to market with our AI background generator and Instant Backgrounds, and this helped us quickly build momentum, and just a year later we’re now the leader in AI photo-editing.”
Lim doesn’t believe AI, nor AI sustainability, is uncharted water for PhotoRoom.
“As a user-centric, mobile-first and AI-first company it has always been a priority for our team. We are composed of experts who have been working on these problems for many years. And we have developed our software with a great degree of consideration towards our energy consumption.”
PhotoRoom also reduces the environmental impact of client workflows, according to Lim.
“One ecommerce business told us they used to fly to the Alps to photograph their products, and now they are generating the same mountain scenes with PhotoRoom in seconds. This has been a game-changer for reducing the carbon footprint of their team, as the air travel industry is one of the biggest polluters on the planet.”
To Join Our WhatsApp Group for the latest Finance related News… Click here to get all the latest and important news..