午夜福利1000集合

ARTIFICIAL intelligence breakthroughs have become a regular occurrence in recent years. One of the most impressive achievements so far was in 2016, when Google DeepMind鈥檚 AlphaGo AI beat champion Lee Sedol at one of the world鈥檚 most complex games, Go.

The feat made headlines around the world as an example of machines besting humans, but in some sense it wasn鈥檛 a fair fight. Sedol鈥檚 brain would have been consuming around 20 watts of power, with only a fraction of that being used for the game itself. By contrast, AlphaGo was using some 5000 watts.

It isn鈥檛 widely appreciated how incredibly energy hungry AI is. If you ran AlphaGo non-stop for a year, the electricity alone would cost about 拢6000. That doesn鈥檛 matter for one-off events, like an epic Go showdown. But it does matter if billions of people want their smartphones to be truly smart, or have their cars drive themselves.

鈥淒eep learning involves performing ever more computations on ever more data鈥�

Many potential uses of AI simply won鈥檛 become widespread if they require too much energy. On the flip side, if the uses are so desirable or profitable that people don鈥檛 care about the costs, it could lead to a surge in electricity consumption and make it even harder to limit further warming of the planet.

AI consumes so much energy because the technique behind these recent breakthroughs, deep learning, involves performing ever more computations on ever more data. 鈥淭he models are getting deeper and getting wider, and they use more and more energy to compute,鈥� says Max Welling of the University of Amsterdam, the Netherlands.

Take image recognition, for instance. This is done by training neural nets that mimic the brain. Images are split up and fed into the first layer of the net, which might look for edges. The results get passed to the next layer, which might look for simple shapes such as a dog鈥檚 tail or a letter, and so on.

The more you want the net to recognise, the more layers it must have and the more complex it has to be. Identifying different breeds of dogs, say, is a much harder task than simply recognising dogs.

The most sophisticated neural nets have grown 鈥渙utrageously large鈥�, as a . Their complexity can be measured by the number of parameters in the net 鈥� and some now have more than 100 billion parameters.

This is important for two reasons, says Welling. First, energy is money. 鈥淭here is an economic ceiling to what is a useful AI technology,鈥� he says. For big online services like Facebook and Google, the cost per use of AI has to be tiny 鈥� in the region of a thousandth of a penny.

Second, mobile devices can only use so much energy. It is not just about battery life, but the limit that comes from ensuring devices don鈥檛 get too hot, Welling says.

This is why voice assistants like Siri or Alexa need an internet connection for full functionality 鈥� your phone or smart speaker doesn鈥檛 have the processing power needed to run the AI locally, or the space to store all the associated data.

All this means that our use of AI can鈥檛 keep expanding indefinitely unless we reduce the energy requirements. 鈥淲hat matters is how much intelligence we can squeeze out per joule,鈥� says Welling.

Battery busters

Take self-driving cars. These require all sorts of extra systems, from cameras to radar, that use power and also increase weight and drag, further increasing energy use. But the single largest consumer of energy besides the engine is the processor running the AI.

According to a study out earlier this year, could use up to 20 per cent more energy than conventional cars. That is a big issue for a battery-powered car, limiting its range and increasing running costs. What鈥檚 more, the study assumes the AI processor consumes about 200 watts, even though current prototypes consume in excess of 2000 watts.

For taxi companies using AI to directly replace human drivers, the savings in wages would probably far outweigh the higher energy costs. But for ordinary car owners this would be a major issue.

Thankfully, people are already working to improve the efficiency of deep-learning algorithms, by getting rid of unnecessary precision in the calculations that greatly increases the processing required, for example.

鈥淓ven if you use fewer bits, you can train a neural network to get the same results,鈥� says Avishek Biswas of the Massachusetts Institute of Technology.

Many big tech companies are also developing specialised hardware for running AI. Nvidia, for instance, has just for self-driving cars 鈥� although the latest version still uses 500 watts.

Google, meanwhile, has created what it calls tensor processing units, or TPUs, designed to run its TensorFlow machine-learning framework, and is renting them out via the cloud.

There are more revolutionary designs in the works, too. Shunting data back and forth between the memory and processor wastes a great of deal of energy, says Biswas. So he has developed a chip intended for smartphones that use by around 95 per cent by carrying out key operations in the memory itself.

The pace of progress is astonishing. The latest version of AlphaGo, called AlphaGo Zero, , and consumes around 400 watts.

But will these new chips actually bring down the overall energy use of AI? As AI becomes more energy efficient, we will build better systems and use it to do more, meaning there may be no net gain. 鈥淐ertainly energy use will keep on going up with more sophisticated AI,鈥� says Welling.

In fact, AI is set to play a major part in the world economy. Global is forecast to reach $60 billion in 2021, according to market research company IDC. Meanwhile, another research company, Gartner, predicts that the value of AI-derived business will .

But we have no idea what this means for global energy use. It is estimated that computing and the internet use about 5 per cent of electricity globally, and AI is clearly a fraction of this at the moment.

鈥淣ot much is known about it at this stage,鈥� says Anders Andrae of Huawei Technologies, who studies the energy consumption of information technology. 鈥淏ut it is very dangerous to say that this is nothing to care about.鈥�

That鈥檚 because what matters is not just how much energy AI itself uses, but also the knock-on effects. To go back to the example of self-driving cars, in theory they could operate in a way that reduces overall energy use despite their higher energy requirements. For instance, they could line up behind each other to reduce drag and communicate with each other so they don鈥檛 need to stop at intersections.

But most of these savings require a high proportion of the cars on the roads to be autonomous 鈥� just as a social network like Facebook can only really take off once it reaches a critical mass of people. 鈥淭he network-dependent effects won鈥檛 happen in the early phases,鈥� says Zia Wadud at the University of Leeds, UK, who studies transport and energy.

鈥淓nergy use will keep going up as we develop more sophisticated artificial intelligence鈥�

People with self-driving cars are likely to commute longer distances, because they won鈥檛 need to focus on the road and can use the time for browsing the internet or watching videos. Wadud thinks this could 鈥� and his study doesn鈥檛 count the AI part, or the massive increase in data traffic these vehicles are expected to generate.

In recent years, electricity demand has remained flat despite soaring internet traffic, thanks to big improvements in efficiency. But Andrae thinks we have reached the limits of the easy gains in efficiency.

If the amount of data we generate rises without big efficiency gains, information and communication technology energy use could soar to 20 per cent of all electricity by 2025 and account for a 20th of all carbon emissions, Andrae warns. 鈥淭his could be a very huge problem,鈥� he says.

High costs make this worst-case scenario unlikely. But Andrae鈥檚 work didn鈥檛 factor in AI and could even be underestimating data growth, for instance from lots of AI cars talking to each other.

鈥淚 think AI will generate even more data than we have seen,鈥� he says. 鈥淚f more data means money, they will use artificial intelligence to make more data, and more data means more electricity.鈥� And until our electricity systems are 100 per cent renewable, more electricity means more global warming.

This article appeared in print under the headline 鈥淜nowledge means power鈥�