Source: Chen Gen Talks about Technology

Image Source: Generated by Wujie AI

Today, AI large models represented by ChatGPT are bringing about huge changes to human society, but they are also controversial due to energy consumption.

The latest article from The Economist states: high-performance computing facilities, including supercomputers, are becoming major energy consumers. According to the International Energy Agency, the electricity consumption of data centers accounts for 1.5% to 2% of global electricity consumption, roughly equivalent to the entire electricity consumption of the UK. It is estimated that by 2030, this proportion will rise to 4%.

Artificial intelligence not only consumes electricity, but also water. Google's 2023 environmental report shows that it consumed 5.6 billion gallons (about 21.2 billion liters) of water in 2022, equivalent to the water used on 37 golf courses. Of this, 5.2 billion gallons were used for the company's data centers, a 20% increase from 2021.

Faced with huge energy costs, the economic viability has become an urgent issue for ChatGPT to move towards the future. And if the energy consumption problem is to be solved, any optimization measures based on existing technology and architecture will be just a stopgap solution. In this context, breakthroughs in cutting-edge technology may be the ultimate solution to solving the energy consumption dilemma of AI.

Artificial intelligence is devouring energy

From the essence of computation, computation is the process of transforming data from disorder to order, which requires a certain amount of energy input.

In terms of quantity alone, according to incomplete statistics, around 5% of global electricity generation in 2020 was used for computing power, and this figure is likely to increase to around 15% to 25% by 2030, meaning that the electricity consumption of the computing industry will be comparable to that of energy-intensive industries such as manufacturing.

In 2020, the electricity consumption of data centers in China exceeded 200 billion kilowatt-hours, which is twice the total electricity generation of the Three Gorges Dam and Gezhouba Dam power plants (approximately 100 billion kilowatt-hours).

In fact, for the computing industry, electricity cost is also the most core cost apart from chip cost.

If this consumed electricity is not generated from renewable energy sources, it will result in carbon emissions. This is also the reason why machine learning models, including ChatGPT, also produce carbon emissions.

Data shows that training GPT-3 consumed 1,287 MWh of electricity, equivalent to emitting 552 tons of carbon. Sustainable data researcher Casper Ludvigsen analyzed, "The large emissions of GPT-3 can be partially explained by the fact that it was trained on older, less efficient hardware, but because there is no standardized method for measuring carbon emissions, these numbers are based on estimates. In addition, the specific amount of carbon emissions that should be allocated to training ChatGPT is also quite vague. It should be noted that since reinforcement learning itself also requires additional electricity consumption, the carbon emissions generated by ChatGPT during the model training phase should be greater than this value." Simply put, the 552 tons of emissions are equivalent to the energy consumption of 126 Danish households per year.

During the operational phase, although the amount of electricity consumed when operating ChatGPT is small, due to the potential occurrence of billions of times globally every day, the cumulative effect may make it the second largest source of carbon emissions.

Databoxer co-founder Chris Boulton explained a calculation method, "First, we estimate that each response word on an A100 GPU takes 0.35 seconds. Assuming there are 1 million users, each with 10 questions, resulting in 10 million responses and 300 million words per day, at 0.35 seconds per word, we can calculate that the A100 GPU runs for 29,167 hours per day."

Cloud Carbon Footprint lists the minimum power consumption of A100 GPUs in Azure data centers as 46W and the maximum as 407W. Since there are likely not many ChatGPT processors idle, using the top end of this range for consumption, the daily power consumption will reach 11,870 kWh.

Chris Boulton stated, "The emission factor in the western United States is 0.000322167 tons/kWh, so it will produce 3.82 tons of carbon dioxide equivalent per day, and the average American emits about 15 tons of carbon dioxide equivalent per year, which is equivalent to the carbon dioxide emission rate of 93 Americans per year."

Although the "virtual" nature makes it easy for people to overlook the carbon footprint of digital products, in fact, the internet has long been one of the largest coal-powered machines on Earth. Research at the University of California, Berkeley on power consumption and artificial intelligence suggests that artificial intelligence is almost devouring energy.

For example, Google's pre-trained language model T5 used 86 megawatts of electricity, resulting in 47 metric tons of carbon dioxide emissions; Google's open-domain chatbot Meena used 232 megawatts of electricity, resulting in 96 metric tons of carbon dioxide emissions; Google's language translation framework GShard used 24 megawatts of electricity, resulting in 4.3 metric tons of carbon dioxide emissions; Google's routing algorithm Switch Transformer used 179 megawatts of electricity, resulting in 59 metric tons of carbon dioxide emissions.

The computing power used in deep learning increased 300,000 times between 2012 and 2018, making GPT-3 appear to have the largest impact on climate. However, when it works alongside the human brain, the energy consumption of the human brain is only 0.002% of that of the machine.

Not only does it consume electricity, it also consumes water

In addition to its astonishing electricity consumption, artificial intelligence also consumes a significant amount of water.

In fact, whether it is electricity or water consumption, it is inseparable from the backbone of the digital world, the data center. As servers and network equipment that power the internet and store large amounts of data, data centers require a large amount of energy to operate, and cooling systems are one of the main drivers of energy consumption.

The truth is, a super-large data center consumes nearly 100 million kilowatt-hours of electricity each year, and the development of generative AI further increases the energy consumption of data centers. Because large models often require tens of thousands of GPUs, the training period can range from a few weeks to several months, requiring a large amount of electricity to support the process.

A large amount of heat is generated during the operation of data center servers, and water cooling is the most common method for servers, which in turn leads to significant water consumption. Data shows that GPT-3 consumed nearly 700 tons of water during training, and then requires 500 milliliters of water for every 20-50 questions answered.

Research at Virginia Tech pointed out that data centers need to consume an average of 401 tons of water per day for cooling, equivalent to the water usage of about 100,000 households. In 2022, Meta used over 2.6 million cubic meters (about 697 million gallons) of water, mainly for data centers. Its latest large language model "Llama 2" also requires a large amount of water for training. Nevertheless, in 2022, one-fifth of Meta's data centers faced "water shortages."

In addition, another important infrastructure for artificial intelligence, chips, also consumes a large amount of energy and water resources in the manufacturing process. In terms of energy, the chip manufacturing process requires a large amount of electricity, especially for advanced process chips. A report by the East Asia branch of the international environmental organization Greenpeace, titled "Electricity Consumption and Carbon Emissions Forecast in the Consumer Electronics Supply Chain," studied the carbon emissions of 13 leading electronic manufacturing companies in East Asia, including Samsung Electronics and TSMC, and stated that carbon emissions in the electronics manufacturing industry, especially the semiconductor industry, are soaring, and by 2030, global electricity consumption in the semiconductor industry will soar to 237 terawatt-hours.

In terms of water consumption, the silicon wafer process requires "ultrapure water" for cleaning, and the higher the chip manufacturing process, the more water is consumed. It takes about 32 kilograms of water to produce a 2-gram computer chip. Producing an 8-inch wafer consumes about 250 tons of water per hour, while a 12-inch wafer can consume up to 500 tons of water.

TSMC has an annual wafer capacity of about 30 million pieces, consuming about 80 million tons of water for chip production. Sufficient water resources have become a necessary condition for the development of the chip industry. In July 2023, the Japanese Ministry of Economy, Trade and Industry decided to establish a new system to provide subsidies for the construction of facilities to supply industrial water to semiconductor factories, to ensure the supply of industrial water needed for semiconductor production.

In the long run, the widespread application of generative AI, autonomous driving, and other applications will further lead to the growth of the chip manufacturing industry, accompanied by the massive consumption of energy resources.

Who can save the energy consumption of AI?

It can be said that today, energy consumption has become a weak point restricting the development of AI. According to the current technological route and development model, the progress of AI will lead to two problems:

On the one hand, the scale of data centers will become increasingly large, and their power consumption will also increase, and their operation will become slower.

Obviously, with the popularization of AI applications, the demand for data center resources by AI will increase sharply. Large-scale data centers require a large amount of electricity to operate servers, storage devices, and cooling systems. This leads to increased energy consumption and may also cause problems with energy supply stability and environmental impact. The continued growth of data centers may also put pressure on energy supply, and the result of relying on traditional energy sources to meet the energy needs of data centers may be rising energy prices and unstable supply. Of course, the high energy consumption of data centers will also have an impact on the environment, including carbon dioxide emissions and energy consumption.

On the other hand, AI chips are evolving towards high computing power and high integration, relying on process technology to support the growth of peak computing power, and the more advanced the process, the greater the power and water consumption.

So, faced with such huge energy consumption of AI, do we have a better way? In fact, the best way to solve technological dilemmas is to develop new technologies.

On the one hand, the progress of AI in the post-Moore era requires finding new and more reliable paradigms and methods.

In fact, today, the huge energy consumption problem caused by artificial intelligence is closely related to the way artificial intelligence achieves intelligence.

We can compare the current construction and operation of artificial neural networks to a group of independent artificial "neurons" working together. Each neuron is like a small computing unit that can receive information, perform some calculations, and then produce output. The current artificial neural network is constructed by cleverly designing the connections of these computing units, and once trained, they can perform specific tasks.

But artificial neural networks also have their limitations. For example, if we need to use a neural network to distinguish between circles and squares. One way is to place two neurons in the output layer, one representing a circle and the other representing a square. However, if we want the neural network to also be able to distinguish the colors of shapes, such as blue and red, we would need four output neurons: blue circle, blue square, red circle, and red square.

In other words, as the complexity of tasks increases, the structure of neural networks also needs more neurons to process more information. The reason for this is that the way artificial neural systems achieve intelligence is not the same as the way the human brain perceives the natural world, but rather "for all combinations, the artificial intelligence neural system must have a corresponding neuron."

In contrast, the human brain can effortlessly accomplish most learning, because the information in the brain is represented by the activity of a large number of neurons. In other words, the perception of a red square in the human brain is not encoded as the activity of a single neuron, but as the activity of thousands of neurons. The same group of neurons, triggered in different ways, can represent a completely different concept.

It can be seen that human brain computation is a completely different way of computing. And if this computing method is applied to artificial intelligence technology, it will significantly reduce the energy consumption of artificial intelligence. This computing method is called "hyperdimensional computing." That is, to mimic the computational method of the human brain, using high-dimensional mathematical space to perform calculations in order to achieve a more efficient and intelligent computing process.

For example, the traditional architectural design pattern is two-dimensional, where we can only draw blueprints on a flat surface, with each blueprint representing different aspects of the building, such as floor layout, wiring, etc. But as buildings become more complex, we need more and more blueprints to represent all the details, which takes up a lot of time and paper.

Hyperdimensional computing is like providing us with a completely new design method. We can design buildings in three-dimensional space, with each dimension representing a property, such as length, width, height, material, color, etc. And we can also design in higher-dimensional space, such as the fourth dimension representing the changes in the building at different points in time. This allows us to complete all the designs on a single super blueprint, without needing a pile of two-dimensional blueprints, greatly improving efficiency.

Similarly, the energy consumption issue in AI training can be compared to architectural design. Traditional deep learning requires a large amount of computing resources to process each feature or property, while hyperdimensional computing processes all features in a unified high-dimensional space. This way, AI only needs to perform calculations once to simultaneously perceive multiple features, saving a lot of computing time and energy consumption.

On the other hand, finding new energy resource solutions, such as nuclear fusion technology, is another approach. Nuclear fusion power generation technology is considered one of the ultimate solutions to global carbon emissions because it produces virtually no nuclear waste during the production process and has no carbon emission pollution.

In May 2023, Microsoft signed a purchase agreement with the nuclear fusion startup company Helion Energy, becoming the company's first customer, and will purchase its electricity when the company completes the world's first nuclear fusion power plant in 2028. In the long run, even if AI achieves a reduction in energy consumption per unit of computing power through hyperdimensional computing, breakthroughs in nuclear fusion technology or other low-carbon energy technologies can still free AI development from carbon emission constraints, providing significant support and impetus for AI development.

In the end, the energy resource consumption issues brought about by technology can only be fundamentally solved at the technological level. Technology constrains the development of technology and also drives the development of technology, as it has always been since ancient times.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。