[ad_1]
The tens of thousands of GPUs used to train large language models (LLMs) are known to consume large amounts of energy, leading to warnings about their potential impact on the Earth’s climate.
However, the infrastructure that supports AI is not the main threat, according to the Center for Data Innovation (CDI), a Washington, D.C.-based think tank backed by tech giants such as Intel, Microsoft, Google, Meta and AMD.
In a recent report [PDF]The center believes that many concerns about AI power consumption are exaggerated and stem from misinterpretations of data. The organization also believes that artificial intelligence may have a positive impact on the earth’s climate by replacing less efficient processes and optimizing others.
“It can be misleading to discuss trends in energy use of AI systems without taking into account the substitution effects of technology. Many digital technologies help decarbonize the economy by replacing moving atoms with moving bits,” the group wrote.
The center’s document points to a study [PDF] Cornell University discovers using AI to write a page of text creates CO2 Emissions are 130 to 1,500 times lower than what Americans would produce doing the same activities using a standard laptop – although that number also includes carbon emissions from living and commuting.However, a closer look at these numbers reveals that they ignore 552 tonnes of CO22 First generated by training ChatGPT.
Arguably, the amount of power used to train LL.M.s pales in comparison to the power consumed by deploying LL.M.s at scale, a process called inference. AWS estimates that inference accounts for 90% of model costs, while Meta accounts for closer to 65%. The model is also retrained from time to time.
The CDI report also shows that just as smart thermostats can reduce a home’s energy consumption and carbon footprint, artificial intelligence can achieve similar efficiencies by predicting grid demand in advance. Other examples include using artificial intelligence to determine how much water or chemical fertilizers farmers should use for optimal efficiency, or tracking methane emissions through satellite data.
Of course, to know whether AI is actually making things better, we need to measure it, and according to CID, there’s plenty of room for improvement in this area.
Why are so many estimates wrong?
According to the Center for Data Innovation, this isn’t the first time technology energy consumption has made sensational headlines.
The group cites a statement from the heyday of the dot-com era that estimated the digital economy would account for half of the grid’s resources within a decade. Decades later, the International Energy Agency (IEA) estimates that data centers and networks account for only 1-1.5% of global energy use.
It’s a cutesy number for the center’s supporters, whose various actions have earned them years of antitrust lawsuits that jeopardize their social license.
But it’s also a number that’s difficult to take at face value because data centers are complex systems. The CDI study points out, without irony, that measuring the carbon footprint or energy consumption of activities such as training or inferring artificial intelligence models is prone to error.
One of the highlighted examples cites a paper from the University of Massachusetts Amherst that estimates the carbon footprint of Google’s BERT natural language processing model.This information was then used to estimate the carbon footprint of training the neural architecture search model, which came up with 626,155 pounds of carbon dioxide2 emission.
The findings were widely published in the media, however, a later study showed that the actual emissions were 88 times smaller than originally thought.
Where estimates are accurate, other factors, such as the mix of renewable energy sources, cooling technology and even the accelerators themselves, mean they can only truly represent workloads at that location and time, the report argues.
The logic is this: if the same model is trained two years later using a newer accelerator, CO2 The emissions associated with that job may look completely different. Therefore, this means that larger models will not necessarily consume more electricity or produce more greenhouse gases as a by-product.
There are several reasons for this, one of which is that AI hardware is getting faster, and another is that the models that make headlines may not always be the most efficient, leaving room for optimization.
From this chart, we see that more modern accelerators, such as Nvidia’s A100 or Google’s TPUv4, have a greater impact on emissions than parameter size. – Click to enlarge
“Researchers continue to experiment with techniques such as pruning, quantization, and distillation to create more compact AI models that are faster and more energy-efficient with minimal loss of accuracy,” the authors write.
The argument of the CID report seems to be that past attempts to extrapolate power consumption or carbon emissions have not dated well, either because they made too many assumptions, were based on flawed measurements, or because they did not take into account hardware or Software speed innovation.
While model optimization has its merits, the report does seem to ignore the fact that Moore’s Law is slowing down, and generational improvements in performance are not expected to result in commensurate gains in energy efficiency.
Increase visibility, avoid regulation and increase spend
The report makes several recommendations on how policymakers should respond to concerns about AI’s energy footprint.
The first involves developing standards for measuring power consumption and carbon emissions associated with AI training and inference workloads. Once these are established, the Center for Data Innovation recommends that policymakers should encourage voluntary reporting.
“Voluntary” seems to be the key word here. While the group says it is not opposed to regulating artificial intelligence, the authors paint a “Catch-22” in which trying to regulate the industry is a lose-lose situation.
“Policymakers rarely consider that their needs will increase the energy requirements of training and using artificial intelligence models. For example, LL.M. debiasing techniques often add greater energy costs during the training and fine-tuning phases,” the report states. Likewise implementing safeguards to check that the LLM does not return harmful output, such as offensive remarks, may incur additional computational costs during inference.”
In other words, trying to enforce protections may make the model more power-hungry; enforcing power limits and risks makes the model less safe.
Not surprisingly, the final recommendations call on governments, including the United States, to invest in artificial intelligence as a way to decarbonize their operations. This includes using artificial intelligence to optimize buildings, transportation and other city-wide systems.
“To accelerate the use of artificial intelligence across government agencies to achieve this goal, the president should sign an executive order directing the Technology Modernization Fund… to make environmental impact one of the priority investment areas for funded projects,” the group wrote.
Of course, this all requires better GPUs and AI accelerators, either purchased outright or rented from cloud providers. That’s good news for tech companies that make and sell the tools needed to run these models.
So it’s no surprise that Nvidia was keen to highlight the report in a recent blog post. Nvidia’s revenue has surged in recent quarters as demand for artificial intelligence hardware reaches fever pitch. ®
[ad_2]
Source link