Calculating the carbon cost of artificial intelligence

If you are looking for interesting topics, Artificial Intelligence (AI) will not disappoint you. Artificial intelligence encompasses a set of powerful, mind-bending statistical algorithms that can play chess, decipher sloppy handwriting, understand speech, classify satellite images, and more. The availability of giant data sets for training machine learning models has been one of the key factors in the success of artificial intelligence. But all this computational work isn't free. Some AI experts are increasingly concerned about the environmental impacts associated with building new algorithms, a debate that has spurred new ideas on how to make machines learn more efficiently to reduce AI's carbon footprint.

Back to Earth

To get into the details, we first need to consider the thousands of data centers (all over the world) that operate 24/7 Handle our calculation requests. For AI researchers, these tasks include training multi-layered algorithms with billions of data elements (or tokens—word bits equivalent to four characters or about 3/4 of a word in English). The computational effort involved is astonishing. AI infrastructure company Lambda has provided some interesting facts about GPT-3, OpenAI’s powerful natural language model for generating human-like text. According to Lambda's analysis, if you wanted to train GPT-3's 175 billion parameter model on a single NVIDIA RTX 8000, it would take 665 years, which is no slouch when it comes to graphics cards.

Simply put, the larger the model, the better the performance. Lambda’s team pointed out that the size of state-of-the-art language models is growing at a rate of 10 times every year, which brings us back to the understanding of AI. Growing concerns about carbon footprint. Back in the data center, it's possible to add more numbers to the discussion, at least at a high level. According to estimates by the International Energy Agency (IEA), the total electricity consumption of data centers worldwide is between 200 and 250 TWh. To make this number easier to visualize, assuming locomotives and rolling stock consume an average of 2.5 kilowatt hours per kilometer traveled, 225 terawatt hours is enough to enable a high-speed electric train to travel 9 million kilometers. While only a portion (in data centers) will be used to train and run AI models, sources indicate that the computing demands for machine learning and training are outpacing the average growth in data center activity.

At this point, it's fair to acknowledge that data centers do a good job of managing their energy needs - environmental concerns are a motivator, but it's worth mentioning that electricity is a significant operating expense, "Mission critical" to every facility. Despite a surge in global internet traffic, up 40% in 2020 alone, data center energy use has remained relatively stable over the past decade. "Strong growth in demand for data center services continues to be offset by continued efficiency gains in servers, storage, network switches and data center infrastructure, as well as the growing share of services catered for by efficient cloud and hyperscale data centres," the IEA wrote.

Photonics and more

Additionally, vertically integrated data center operators such as Amazon, Google, Facebook and others will soon add that their facilities are powered by renewable energy. Naturally, this reduces the environmental burden of data processing, as the electricity to power racks of computing hardware and necessary ancillary services such as heating, cooling and lighting can come from the sun and wind. However, as the Financial Times has not chosen, even if a data center energy agreement may offset 100% of its electricity consumption through renewable energy, the facility may still consume fossil fuels when wind and solar are not available. There is also the need to consider the embedded carbon emissions of the computing device itself, as manufacturing methods and component material sourcing activities also create carbon emissions - something Microsoft acknowledges.

Earlier this year, Microsoft discussed the topic of efficient model training in a recent blog post. Developers are busy exploring ways to shrink AI’s carbon footprint—or at least curb its growth. Steps here include looking for ways to reduce computational and memory requirements during model fine-tuning, with recommendations recommending a threefold reduction in GPU usage during this stage of the process. Model compression also shows promise, in which sub-layers of AI data are pruned into sparser but still representative versions of previously combined conditions. Here, research shows that compressing models may require about 40% less training time while achieving similar algorithmic results.

Developers can also benefit from monitoring tools that will pinpoint gains made by optimizing code or data hosting arrangements. “CodeCarbon is a lightweight software package that integrates seamlessly into your Python code base,” write the inventors, who make their tool available for free. "It estimates the amount of carbon dioxide (CO2) produced by cloud or personal computing resources used to execute code."

Full Circle

The cycle continues, and more energy-efficient AI may be deployed in the future to help guide more efficient data center operations to reduce—you guessed it—AI’s carbon footprint . Today, Cirrus Nexus provides available software that data center operators can use to assign a cost to carbon and propagate it through artificial intelligence algorithms. The results not only show CO2 calculations, but also provide insights into the ways in which users can configure their facilities to maximize available environmental benefits.

Making the carbon footprint of the algorithms powering today’s technology visible helps in multiple ways. It’s opening up discussions about the most effective ways to train future artificial intelligence, making IT departments and their customers more accountable for the environmental costs of computing. In the end, it could be good for business. Notably, Amazon released a customer carbon footprint tool earlier this year, and other major companies like Google allow customers to export cloud carbon emissions information — a service currently in preview.

The above is the detailed content of Calculating the carbon cost of artificial intelligence. For more information, please follow other related articles on the PHP Chinese website!