The U.S. Energy Act of 2020 required the Department of Energy to update the 2016 Lawrence Berkeley National Laboratory (LBNL) study.US Data Center Energy Usage ReportFor Data Center World 2024, event organizers asked the authors of the report to provide an update on progress and clarify initial findings. This report, when published, should provide interesting comparisons with the 2016 study.
Armand ShehabiLBNL's Energy/Environmental Policy Staff Scientists to compile a new report based on the rise of purpose-built hardware, artificial intelligence, edge computing and rapidly growing data demands in an increasingly interconnected world. presented the assumptions used.
He highlighted the advances that data centers have made over the past two decades compared to the first few years of the 2000s. From 2000 to 2005, US data center power usage doubled. Most of this consumption was due to volume server and data center cooling/power.
“Most forecasts at the time predicted that data center power usage would continue to grow rapidly,” Shehabi said. “We believed that eventually we would run out of power.”
Fortunately, consumption has leveled off. From 2010 to 2018, data center electricity usage increased by only 6% globally, despite significant increases in computing and storage.
“We've increased our compute instances by 550%, but we're much more efficient,” Shehabi said.
Data center modeling in 2020 predicted that additional energy efficiency efforts could potentially double compute demands. However, this could not have predicted the AI craze and the deployment of advanced GPUs. Researchers are now rushing to complete a major update to this year's report, taking these factors into account.
New research for a rapidly changing world
Shehabi works closely with Sarah Smith, Energy and Environment Policy researcher at LBNL. She compiles the data and adds factors such as cryptocurrencies, carbon consumption, water usage, and other parameters to the analysis. Smith also lends her wisdom to Thermal Design Power (TDP) and how to do it. Power ratings and maximum power may vary depending on server type.
“We assumed that the maximum operating power of the AI server was equal to TDP, but we believe that the model may be more accurate if the maximum power is calculated at 80% of TDP,” says Data Center World's Smith. ”.
Additionally, researchers are trying to estimate the average server usage for the entire year, taking into account downtime. The current assumption is 70% utilization, but the uncertainty factor ranges from 30% to 80%. More work needs to be done to define this correctly to increase the accuracy of the model.
“Average server utilization varies by use case and data center type,” Shehabi said. “We are trying to narrow the range of uncertainty.”
AI makes modeling and prediction even more difficult. But despite the wide range of uncertainties, everyone agrees that electricity consumption will skyrocket between 2024 and 2030. The power consumption of AI will far exceed that of traditional servers.
“Storage and network power consumption is also increasing, but not nearly as much as CPU and GPU utilization,” Shehabi said.
Cooling system prerequisites
The 2024 report takes into account the effects of liquid and other forms of cooling. This is an even more uncertain area for researchers.
“AI is driving the use of liquid cooling,” Shebabi says. “However, due to water scarcity and sustainability concerns, large data centers are moving from chilled water systems to direct expansion systems.”
He asked industry groups and owners for help in understanding exactly where all data centers in the United States are located, especially those of hyperscalers and cryptocurrency miners. Additionally, Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE) estimates are provided for small to medium sized, hyperscalar, and liquid-cooled AI data centers.
Although all areas have not yet been accurately calculated and compiled, power reserve estimates for US data centers are in line with expectations, with sharp increases in power usage across the board.
“We don't yet know how much AI will increase electricity demand, but we do know that it will increase rapidly,” Shehabi said.
Data gaps remain
As mentioned above, many data gaps remain that impede analysis. Other areas of ongoing research include how much power data centers use from the grid and how much power they use from other on-site power sources.
Still, excitement is already rising over the findings revealed in the final report. With the advent of AI, planners can see how AI will impact the grid, how the grid will react, and how quickly efficiency and technological advances can keep up with growth. It will look like this. This will give researchers and stakeholders insight into how current efforts to drive the transition to EVs and the electrification of buildings and other industries will impact the power grid.
“Historically, utilities have often overestimated capacity demand,” Shehabi said. “We want to see if that applies to AI as well.”
Some believe that the United States has enough power now available or under development to cover the increase due to AI. The problem is that power isn't being delivered where it's needed.
“We expect power shortages to be a regional rather than a national issue,” Smith said.
She pointed to opportunities for data centers in demand response. Many data centers have generators and other forms of backup power that are rarely used. In some regions, you may be willing to pay a data center fee to have your backup generation source on standby. When the grid needs more power, it notifies the data center to switch to backup generation. Data centers are paid year-round for services they provide only a few hours here and there.
Streamlining AI
What this report may not have considered is the impact of innovation on AI power usage. The software has been streamlined to better work with GPUs. More integrated AI infrastructure and interconnections are being created that facilitate AI while consuming less power.
Michael Azoff, Chief Analyst Omdia noted that companies are likely to gravitate towards smaller, more focused models rather than developing huge models like ChatGPT.
“ChatGPT has 1.8 trillion data points, but there are smaller models emerging with perhaps 2.5 billion data points that can provide better results for enterprises,” Azoff said. “One company has successfully built such models using CPUs rather than GPUs, significantly reducing processing requirements.”