AI raises concerns about environmental destruction like Bitcoin
Bitcoin and cryptocurrencies have been criticized for being environmentally destructive due to their high electricity consumption, while AI can consume even more energy.
According to its second quarter 2023 financial report, Nvidia said revenue from chips used in data centers, mainly for AI training, has doubled. According to Bloomberg , this shows the need for generative AI models like ChatGPT may not have peaked yet.
However, this trend causes concerns about energy consumption from data centers when Nvidia's new chip model, GH200 Grace Hopper Superchip, is only the size of a postcard but consumes up to 1,000 watts of electricity. equivalent to a portable heater. Meanwhile, super AI training machines have many such chips connected to each other.
Both Bitcoin mining and AI training require a lot of energy. Photo: CBJ
Research published this month by HuggingFace shows that with 176 billion parameters from 1.6 TB of data, a cluster of 384 Nvidia A100 graphics processors would take more than 118 days to analyze. It is estimated that the system will generate from 24.7 tons to 50.5 tons of CO2 during operation.
However, that is only the first training step of a super AI model. Amazon statistics show that 90% of the costs from operating an artificial intelligence system will lie in the next stage, when users query the model to receive results, for example asking ChatGPT to provide a recipe. cake. At this stage, training an AI system can generate 10 times more CO2 than the first stage, or about 500 tons, equivalent to driving a car one million miles or 500 flights from New York to Frankfurt (more than 6,200 kilometer).
According to research by The Shift Network published at the end of May, information technology, computing and data processing systems account for 3.7% of global carbon emissions, higher than emissions from the industry. aviation with 2.4%. It is forecast that the computing sector will exceed 14% of global emissions by 2040.
Previously, experts at the University of Colorado Riverside and the University of Texas estimated that a basic conversation with ChatGPT , with 20-50 questions and responses, could consume 500 ml of water. . The total amount of water will increase to a huge level due to the fact that the number of users is up to hundreds of millions and billions of questions have been asked over time.
"Large language models and machine learning contribute a huge carbon footprint to the environment," said Philipp von Bieberstein, co-founder of carbon-tracking software firm Climatiq.
Also in May, an estimate from CarbonCredits found that an AI search query generates four times more carbon emissions than a Google search. In the future, when AI is enhanced, this index may change.
Meanwhile, according to research from Cambridge University last year, Bitcoin mining generates 72.5 million tons of CO2 per year. Currently, the emissions caused by Bitcoin mining are larger than those of AI. However, in the context of a series of large technology companies racing to build artificial intelligence products such as OpenAI, Amazon, Microsoft, Google, Meta from the US or Baidu, Tencent, Alibaba from China, future emissions are uncertain. will increase many times.
"In the race to surpass each other, companies will not stand still. They will continue to spend billions of dollars buying energy-consuming processors for increasingly large data analysis systems," Bloomberg commented . .
Besides, Bitcoin mining activities have gradually switched to using green energy such as hydroelectric power, wind power or solar power, most AI training machines still use electricity from fossil fuels. According to research on Newscientist earlier this year, less than 25% of AI systems use low-carbon energy sources such as hydroelectricity and nuclear power during operation. The reason is that AI systems are often based in city centers due to the need for high Internet speeds, thousands of miles away from hydroelectric dams or solar panel arrays.
Experts say that although it consumes a lot of energy, AI is not as negative as the digital currency mining model. Most types of Bitcoin mining currently do not bring much value to the world, while innovative artificial intelligence does not have this problem.