The growing interest in AI systems will add more pressure on global power grids with the potential to compete with Bitcoin’s huge energy consumption. Fortunately, the major cryptocurrency has shown us a way to mitigate the impact.
Nvidia Corp.’s data center revenue doubling last quarter shows that demand for generative applications such as ChatGPT has not yet peaked. The US chipmaker is the main supplier of code in this AI gold rush, but these processors are neither cheap nor efficient. Their latest flagship product, the postcard-sized GH200 Grace Hopper Superchip, consumes up to 1,000 watts, the equivalent of a space heater.
Although most customers would opt for something a little less fancy than the Superchip, buy them in bulk to connect them to a massive AI server, and that’s where the hunger for electricity really begins. A study published last year looked at the required energy intake. To train a single large language model that is used to generate text in multiple languages.
BLOOM from startup HuggingFace used 176 billion parameters from 1.6 terabytes of data. According to the study authors, it took more than 118 days to run a cluster of 384 Nvidia A100 graphics processors. They estimated that the electricity consumption from running many GPUs for an extended period might generate 24.7 metric tons (54,000 lb) of carbon dioxide. But the actual cost doubles to 50.5 tons if network connections and downtime of the entire system are taken into account.
Until then, model training is just the beginning. According to an estimate by Amazon.com, which runs its own AI servers, 90% of the spending generated by running the AI occurs in the next stage, when users query the form to get results, such as asking it for ChatGPT chocolate cake. recipes. . It is difficult to calculate the energy consumption required to execute the data, called inference, but it is believed to be around 10 times what is needed for the first training phase, which is 500 tons of carbon dioxide. A single AI query could have a carbon footprint four times larger than a Google search, according to one estimate.
Brute-force number manipulation is built into Bitcoin’s design and helps explain why a wave of semiconductors and servers spread around the world hoping to mine digital gold. An ongoing study at the University of Cambridge estimates that bitcoin is responsible for 72.5 million tons of carbon dioxide. This number could reach 3 million tons if all bitcoin mines were powered by hydroelectric power. Compared to the waste of cryptocurrency, 500 tons of CO2 from a single round of training and deployment seems like nothing. However, it is still the equivalent of driving a million miles in a gas car or 500 flights from New York to Frankfurt.
And it’s still early. At least a dozen major technology companies are rushing to build and deploy generative AI products, including Amazon.com, Alphabet Inc., Microsoft Corp., OpenAI, Meta Platforms Inc., Baidu Inc., Tencent Holdings Ltd., and Alibaba Group Holding. . Ltd. Since everyone is in a race to outdo each other, they will not sit still once the model is trained; They will continue to buy power-hungry processors to analyze ever-increasing amounts of data. Once this is done, they will compete with each other to present the results to consumers in the form of college articles, fake videos, and synthetic versions of Pink Floyd’s music.
This is made worse by the fact that most AI training currently runs on fossil fuels. These server farms quickly expanded into existing locations, which are often thousands of miles away from hydroelectric dams or solar panels. Since network latency is an issue when responding to Internet requests, it needs to be close to the end user and not thousands of miles away.
But bitcoin has already made a way for the AI industry to follow. The cool climate combined with abundant renewable energy has made it the ideal location for an energy-hungry cryptocurrency miner, and Iceland’s arctic air and abundant thermal energy make the country an ideal choice. China may also find new use for the many hydroelectric power plants that attracted mining rigs but lost business after Beijing cracked down on the digital currency. Foreign AI providers won’t have access to it, of course, but Chinese tech giants know they have this resource at their fingertips as their energy needs grow.
There is another advantage of the AI bitcoin exchange in these server farms. While the cryptocurrency has attracted many speculators and billions of dollars in investments, it still fails to add much value to the world. Generative AI does not have this problem; Just ask ChatGPT.
This memo does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.
“Beeraholic. Friend of animals everywhere. Evil web scholar. Zombie maven.”