Energy consumption when training LLMs in 2022 (in MWh)

Energy consumption of artificial intelligence (AI) models in training is considerable, with both GPT-3, the original release of the current iteration of OpenAI's popular ChatGPT, and Gopher consuming well over a thousand-megawatt hours of energy simply for training. As this is only for the training model it is likely that the energy consumption for the entire usage and lifetime of GPT-3 and other large language models (LLMs) is significantly higher. The largest consumer of energy, GPT-3, consumed roughly the equivalent of 200 Germans in 2022. While not a staggering amount, it is a considerable use of energy. Energy savings through AI While it is undoubtedly true that training LLMs takes a considerable amount of energy, the energy savings are also likely to be substantial. Any AI model that improves processes by minute numbers might save hours on shipment, liters of fuel, or dozens of computations. Each one of these uses energy as well and the sum of energy saved through a LLM might vastly outperform its energy cost. A good example is mobile phone operators, of which a third expect that AI might reduce power consumption by ten to fifteen percent. Considering that much of the world uses mobile phones this would be a considerable energy saver. Emissions are considerable The amount of CO2 emissions from training LLMs is also considerable, with GPT-3 producing nearly 500 tonnes of CO2. This again could be radically changed based on the types of energy production creating the emissions. Most data center operators for instance would prefer to have nuclear energy play a key role, a significantly low-emission energy producer.

人工智能（AI）模型在训练过程中所消耗的能源量是相当可观的，以OpenAI流行聊天机器人ChatGPT当前版本的GPT-3及其前身Gopher为例，它们仅训练阶段就耗费了超过一千兆瓦时的能源。鉴于这仅仅是模型训练所需的能源，GPT-3及其他大语言模型（LLMs）在整个使用和生命周期中的能源消耗很可能更为巨大。2022年，GPT-3的能源消耗量相当于约200名德国人的能源消耗量。虽然这一数字并非令人震惊，但能源的使用量无疑是相当可观的。

通过人工智能实现节能尽管训练大语言模型无疑需要大量能源，但节能的潜力也同样巨大。任何通过微小程度提升流程的AI模型都可能节省数小时的运输时间、数升燃料或数十次计算。这些应用同样消耗能源，而通过大语言模型节约的能源总量可能会远远超过其能源成本。以移动运营商为例，其中三分之一预计AI可能将电力消耗降低百分之十至十五。考虑到全球大量使用手机，这将是一个巨大的节能举措。

碳排放量同样可观，训练大语言模型所产生的二氧化碳（CO2）排放量也是相当可观的，GPT-3的排放量接近500吨。这再次表明，根据产生排放的能源生产类型，这一数字可能会有根本性的变化。例如，大多数数据中心运营商更希望核能发挥关键作用，因为核能是一种显著低排放的能源生产者。