laitimes

How long is the power consumption of "AI" enough for you to blow the air conditioner?

How long is the power consumption of "AI" enough for you to blow the air conditioner?

How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?
How long is the power consumption of "AI" enough for you to blow the air conditioner?

Discussing the high power consumption of AI is not a matter of calculation, but of an out-and-out energy problem.

According to the International Energy Agency's (IEA) Electricity 2024 – Analysis and Forecast to 2026, the data center, artificial intelligence (AI) and cryptocurrency industries accounted for 2% of global electricity consumption in 2022 and are expected to double to more than 1,000 TWh by 2026.

Cheng Daning, an assistant researcher at the Institute of Computing Technology of the Chinese Academy of Sciences, said in an interview with The Paper that according to the survey, computing power will develop towards a larger scale, faster interconnection and smaller in the future. There are two reasons for this: on the one hand, the scale of the model is getting larger and larger, which requires greater computing power and larger video memory to support, and on the other hand, the intelligence of edge devices requires lower power consumption, faster deployment, and a large number of edge devices, so the demand is also large. Intelligent devices have become an important part of life, and inference will be a greater demand for computing power in the future.

Cheng Daning also mentioned that from the perspective of industry data, electricity accounts for more than 50% of the total cost of the operating costs of domestic first-line large models.

If AI continues to consume so much power, when will ordinary people be able to use cheap AI? You know, GPT-4 costs $20 a month, and Wenxin Yiyan also costs about 60 yuan a month. More AI tools are starting to opt for pay-as-you-go. There have been reports that Stable Diffusion, a startup that has made a name for itself with AI painting, can no longer afford sky-high computing power bills.

In terms of total volume, it is inevitable that AI will continue to consume more electricity, but technology companies are also looking for ways to reduce the impact of AI.

One way is to reduce its appetite, and it can improve efficiency through optimization algorithms, quantizing and pruning the model are ways to reduce the amount of computation, so it can also reduce power consumption. On the other hand, tech companies are also betting on new energy sources, with OpenAI's CEO, Sam Altman, investing in a fusion startup. Nvidia CEO Jensen Huang is more optimistic, arguing that if computing power advances fast enough, there is no need to worry too much about energy consumption.

Strategies for recovering heat energy are also important. According to the estimation of the computing center of Shanghai Jiao Tong University, the recovery of heat energy can reduce the carbon emissions of the computing center by 4,850 tons per year, of which waste heat recovery accounts for 950 tons. The use case of UCD Ulanqab data center shows that the use of computing power and thermal energy can save about 50% of electricity costs.

As the saying goes, you always have to pay it back when you come out to mix. AI eats so much electricity to cultivate itself, in addition to spitting out heat, can it also pay us a new energy future? Time will tell.

For now, at least, before chatting with an AI, weigh up its exaggerated power consumption.

Resources:

"The "Electricity Maniac" Hidden Behind AI - Snow Leopard Finance Club:

Hatpas://vv.36KR.com/P/2725525783520512

"Why does AI consume so much power when it uses 240 million kWh of electricity for 1 training session?" - Back to Basics:

https://mp.weixin.qq.com/s/VPnzuzUWSOw2coA0uz7ETQ

"A Sentence That Huang Didn't Say, It's on Fire" - 36 Krypton:

Hatpas://36KR.com/P/2687714777493126

"The End of Computing Power Is Actually Electricity?" - Tencent Research Institute:

Hatps://m.huksio.com/article/2861503.html

Read on