Can we make artificial intelligence less attractive in power? These researchers are working on this.

For tests, his team used settings with A100 and H100 graphics processing units from NVIDIA, which are commonly used in data centers today, measuring the amount of energy they used to operate the different large language models (LLMS), spread models that create images or videos based on text inputs, and many other systems of artificial intelligence systems.

The largest LLMA LLAMA LLAMA 3.1 405b, the chat -based Amnesty International with 405 billion teachers. 3352.92 Joules consumed power for every request that works on H100 graphics processing units. This is about 0.93 watts of an hour-less than 2.9 watts watch an hour adapted for the ChatgPT information. These measurements confirmed the improvements in the energy efficiency of the devices. The MIXTRAL 8X22B was the largest LLM, which enabled the team to run on both Apere and Hopper platforms. The form of the model on two Aprice graphics processing units led to 0.32 watts an hour per order, compared to only 0.15 w

However, what remains known is the performance of royal models such as GPT-4, Gemini or Grok. The ML Energy initiative team says it is very difficult for the research community to start reaching solutions to energy efficiency problems when we do not even know exactly what we face. We can make estimates, but Chung insists that they need to be accompanied by a mistake -related analysis. We have nothing like that today.

The most urgent issue, according to Chong and Chuoderi, is a lack of transparency. “Includes or Open AI do not have any incentive to talk about energy consumption. If there is anything, then the launch of the actual numbers will harm them,” said Chaudhry. “But people must understand what is really happening, so we may have somehow to convince them to issue some of these numbers.”

Where the rubber meets the road

“Energy efficiency in data centers similar to the More Law – it only works very widely, instead of one chip,” said Harris Nafidia. He said that energy consumption for each shelf, a unit used in data centers ranging from 10 and 14 NVIDIA GPUS, is rising, but the performance of each watt is improving.