Power consumption and water consumption, who can save AI energy consumption?

Original source: Chen Gen talks about technology

Image source: Generated by Unbounded AI‌

Today, while the large AI model represented by ChatGPT has brought great changes to human society, it is also controversial because of energy consumption.

The Economist's latest publication said: High-performance computing facilities, including supercomputers, are becoming major energy consumers. **According to the International Energy Agency, data centers account for 1.5% to 2% of global electricity consumption, roughly equivalent to the electricity consumption of the entire UK economy. **This is expected to rise to 4% by 2030.

**Artificial intelligence consumes not only electricity, but also water. ** According to the 2023 environmental report released by Google, it will consume 5.6 billion gallons (about 21.2 billion liters) of water in 2022, which is equivalent to the water of 37 golf courses. Of that, 5.2 billion gallons went to the company's data centers, a 20% increase over 2021.

In the face of huge energy consumption costs, artificial intelligence (AI) wants to move towards the future, and economy has become a real problem that ChatGPT needs to solve urgently. And if the energy consumption problem is to be solved, any optimization measures based on the existing technology and architecture will stop the boiling water. In this context, the breakthrough of cutting-edge technology may be the ultimate solution to the AI energy consumption dilemma. .

AI is eating energy

From the essence of computing, computing is the process of changing data from disorder to order, and this process requires a certain amount of energy input.

From the perspective of quantity alone, according to incomplete statistics, about 5% of global power generation in 2020 will be used for computing power consumption, and this figure may increase to about 15% to 25% by 2030, that is, It is said that the proportion of electricity consumption in the computing industry will be on par with large energy-consuming industries such as industry.

In 2020, the power consumption of China's data centers will exceed 200 billion kWh, which is twice the combined power generation of the Three Gorges Dam and Gezhouba Power Plant (about 100 billion kWh).

In fact, for the computing industry, the cost of electricity is also the core cost besides the chip cost.

**If the electricity consumed is not generated by renewable energy, then there will be carbon emissions. This is why machine learning models, also generate carbon emissions. ChatGPT is no exception. **

Data show that training GPT-3 consumes 1287MWh (megawatt hours), which is equivalent to emitting 552 tons of carbon. In this regard, sustainable data researcher Caspar Ludwigsen also analyzed: "GPT-3's large emissions can partly be explained by the fact that it was trained on older, less efficient hardware, but because There is no standardized way to measure CO2 emissions, these numbers are based on estimates. In addition, the standard of how much of this part of the carbon emission value should be allocated to training ChatGPT is also relatively vague. It should be noted that because reinforcement learning itself requires additional It consumes electricity, so the carbon emissions generated by ChatGPT during the model training phase should be greater than this value.” Calculated only by 552 tons of emissions, these are equivalent to the annual energy consumption of 126 Danish households.

**In the operation phase, although the power consumption of people's actions when operating ChatGPT is very small, it may also become the second largest source of carbon emissions due to the fact that there may be one billion times a day in the world. **

Databoxer co-founder Chris Bolton explained one calculation method, "First, we estimate that each response word takes 0.35 seconds on the A100 GPU, assuming 1 million users with 10 questions per user, generating 1000 10,000 responses and 300 million words per day, each word is 0.35 seconds, it can be calculated that the A100 GPU runs for 29167 hours per day."

Cloud Carbon Footprint lists the minimum power consumption of the A100 GPU in the Azure data center of 46W and the maximum power consumption of 407W. Since it is likely that not many ChatGPT processors are idle, the daily power consumption will reach 11870kWh at the top of the range.

Chris Bolton said: "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. The average American is about 15 tons of carbon dioxide equivalent per year. In other words, this is equivalent to the annual carbon dioxide emissions of 93 Americans. The rate is comparable.”

Although the "virtual" attribute makes it easy for people to ignore the carbon books of digital products, in fact, the Internet has long been one of the largest coal-powered machines on the planet. **A Berkeley study on the topic of power consumption and artificial intelligence argues that artificial intelligence nearly eats energy. **

For example, Google’s pre-trained language model T5 used 86 megawatts of electricity and generated 47 metric tons of CO2 emissions; Google’s multi-round open-field chatbot Meena used 232 megawatts of electricity and generated 96 metric tons of CO2 emissions ; The language translation framework developed by Google - GShard used 24 megawatts of electricity and produced 4.3 metric tons of carbon dioxide emissions; the routing algorithm Switch Transformer developed by Google used 179 megawatts of electricity and produced 59 metric tons of carbon dioxide emissions.

Computing power used in deep learning has grown by a factor of 300,000 between 2012 and 2018, making GPT-3 look like the one with the biggest climate impact. However, when it works simultaneously with the human brain, the energy consumption of the human brain is only 0.002% of that of the machine.

Not only consumes electricity, but also consumes water

In addition to staggering power consumption, artificial intelligence is also very water-consuming.

In fact, no matter it is power consumption or water consumption, it is inseparable from the digital center, the pillar of the digital world. As servers and networking equipment that power the internet and store vast amounts of data, global data centers require a lot of energy to operate, and cooling systems are one of the main drivers of energy consumption. **

The truth is that a super-large data center consumes nearly 100 million kilowatt-hours of electricity every year, and the development of generative AI has further increased the energy consumption of the data center. Because large-scale models often require tens of thousands of GPUs, the training period can range from a few weeks to several months, and a large amount of power is required in the process.

Data center servers generate a lot of heat energy during operation, and water cooling is the most common method for servers, which in turn leads to huge water power consumption. Data show that GPT-3 consumes nearly 700 tons of water during training, and then consumes 500 milliliters of water for every 20-50 questions answered.

According to a study by Virginia Tech, data centers consume an average of 401 tons of water per day for cooling, which is equivalent to the water consumption of 100,000 households. Meta used more than 2.6 million cubic meters (about 697 million gallons) of water in 2022, mostly for data centers. Its latest large-scale language model, "Llama 2," also requires a lot of water to train. Even so, in 2022, one fifth of Meta's data centers will experience "water shortage".

In addition, another important infrastructure chip for artificial intelligence, its manufacturing process is also a process that consumes a lot of energy and water resources. In terms of energy, the chip manufacturing process requires a lot of power, especially advanced process chips. International environmental organization Greenpeace East Asia Division's "Consumer Electronics Supply Chain Power Consumption and Carbon Emissions Forecast" report studied the carbon emissions of 13 leading electronics manufacturing companies in East Asia, including Samsung Electronics and TSMC, and said that the electronics manufacturing industry, especially the semiconductor industry Carbon emissions are soaring, and the electricity consumption of the global semiconductor industry will soar to 237 terawatt hours by 2030.

In terms of water resource consumption, the silicon wafer process requires "ultra-pure water" cleaning, and the higher the chip manufacturing process, the more water consumption. It takes about 32 kilograms of water to produce a 2-gram computer chip. Manufacturing 8-inch wafers consumes about 250 tons of water per hour, and 12-inch wafers can reach 500 tons.

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

In the long run, the promotion and application of generative AI and unmanned driving will lead to further growth of the chip manufacturing industry, followed by a large consumption of energy resources.

**Who can save AI from energy consumption? **

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

**On the one hand, the scale of the data center will become larger and larger, and its power consumption will also increase accordingly, and the operation will become slower and slower. **

Obviously, with the popularization of AI applications, AI's demand for data center resources will increase dramatically. Large-scale data centers require large amounts of electricity to run servers, storage devices, and cooling systems. This leads to increased energy consumption, while also raising issues of energy supply stability and environmental impact. The continued growth of data centers may also put pressure on energy supplies, and reliance on traditional energy sources to meet data center energy needs may result in rising energy prices and supply instability. Of course, the high energy consumption of data centers also has environmental impacts, including CO2 emissions and energy consumption.

**On the other hand, AI chips are evolving toward high computing power and high integration, relying on manufacturing processes to support the growth of peak computing power. As manufacturing processes become more advanced, their power consumption and water consumption are also increasing. **

So, in the face of such a huge AI energy consumption, do we have a better way? In fact, the best way to solve the technical dilemma is to develop new technologies.

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

In fact, today, the reason why artificial intelligence will bring huge energy consumption problems is closely related to the way artificial intelligence realizes intelligence.

We can compare the construction and operation of artificial neural networks at this stage to a group of independent artificial "neurons" working together. Each neuron is like a small computational unit that takes in information, does some calculations, and produces an output. Today's artificial neural networks are built by cleverly designing how these computational units are connected so that, 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 a circle from a square. One approach is to place two neurons in the output layer, one for the circle and one for the square. However, if we want the neural network to also be able to distinguish the color of shapes, such as blue and red, then we need four output neurons: blue circle, blue square, red circle, and red square.

That is to say, as the complexity of the task increases, the structure of the neural network also requires more neurons to process more information. The reason is that the way artificial neural networks achieve intelligence is not the way the human brain perceives the natural world, but "for all combinations, the artificial intelligence nervous system must have a corresponding neuron."

By contrast, the human brain can do most of the learning effortlessly, because the information in the brain is represented by the activity of a large number of neurons. That is to say, the human brain's perception of the red square is not encoded as the activity of a single neuron, but is encoded as the activity of thousands of neurons. The same set of neurons, firing in different ways, can represent a completely different concept.

**As you can see, human brain computing is a completely different computing method. And if this calculation method is applied to artificial intelligence technology, it will greatly reduce the energy consumption of artificial intelligence. **And this calculation method is the so-called "super-dimensional calculation". That is to imitate the calculation method of the human brain, and use the high-dimensional mathematical space to perform calculations to achieve a more efficient and intelligent calculation process.

For example, the traditional architectural design model is two-dimensional. We can only draw drawings on a plane. Each drawing represents a different aspect of the building, such as floor layout and wire routing. But as buildings become more and more complex, we need more and more drawings to represent all the details, which takes up a lot of time and paper.

And hyperdimensional computing seems to provide us with a new design method. We can design buildings in three dimensions, with each dimension representing a property such as length, width, height, material, color, etc. Moreover, we can also design in a higher-dimensional space, such as the fourth dimension representing the changes of buildings at different points in time. This allows us to complete all the designs on one super drawing, eliminating the need for a bunch of 2D drawings, which greatly improves efficiency.

Likewise, energy consumption in AI training can be compared to building design. Traditional deep learning requires a large amount of computing resources to process each feature or attribute, while hyperdimensional computing puts all features in a high-dimensional space for processing. In this way, AI can perceive multiple features at the same time only by performing calculations once, thus saving a lot of calculation time and energy consumption.

** On the other hand, find new energy resource solutions, for example, nuclear fusion technology. **Nuclear fusion power generation technology is considered to be one of the ultimate solutions to the global carbon emission problem because it basically does not generate nuclear waste and has no carbon emission pollution during the production process.

In May 2023, Microsoft signed a purchase agreement with Helion Energy, a nuclear fusion start-up company, becoming the company's first customer and will purchase its electricity when the company builds the world's first nuclear fusion power plant in 2028. Moreover, in the long run, even if AI achieves a reduction in energy consumption per unit of computing power through ultra-dimensional computing lamps, breakthroughs in nuclear fusion technology or other low-carbon energy technologies can still make AI development no longer restricted by carbon emissions. Still has great support and promotion significance.

After all, the problem of energy and resource consumption brought about by technology can still only be fundamentally solved at the technical level. Technology restricts the development of technology and also promotes the development of technology, which has been the case since ancient times.