Ms. Rajini Wijayawardana is a 5th-year PhD student in Computer Science at the University of Chicago in the Large-Scale Sustainable Systems group. She primarily works in the lab of Dr. Andrew A. Chien. Her research focuses on sustainable computing, particularly on reducing the environmental impact of data centers and their interactions with power grids. 

​

Wijayawardana emphasized that much of the information about the environmental impacts of computing is unknown because it is still a relatively new and rapidly growing industry. Similar to other interviewees, she mentioned the importance of having transparent sustainability reports from major technology companies. She cited Google as a more transparent company, and she explained that Amazon/AWS provides very limited data. While companies can provide data, they are often unclear about their recording methods, making their significance questionable. There is often a lack of information to understand it properly, and it is not standardized across companies. One way companies can misrepresent their data is by using net carbon accounting, where they cancel out carbon emissions with renewable energy generation. It is further complicated when renewable energy is being sold directly to tech companies instead of to the public power grid. 

​

Her paper, “Reducing the Carbon Impact of Generative AI Inference (today and in 2035),” proposes geographical workload shifting. This paper focuses on inference rather than training and suggests that tasks can be run in regions with clean energy sources to reduce carbon emissions. For example, sending inference requests from Chicago to Canada’s hydroelectric grid. 

Then, Wijayawardana spoke to me about her paper, “Exploding AI Power Use: an Opportunity to Rethink Grid Planning and Management.” This paper proposes a solution to help with the strain AI causes on power grid infrastructure. It takes a very long time to plan for infrastructure that can accommodate a large data center (DC) load, but AI has expanded rapidly. The proposal is to relax power reliability requirements and safely attach more DC load without increasing the loss of load expectation (LOLE). The idea is to accept temporary reductions in power availability in order to allow for data center capacity without destabilizing the grid.

Wijayawardana also discussed the types of data centers. She explains that in many cases, small localized data centers near renewable energy sources can be more effective than large isolated ones. She used the example of Lancium, originally developed to pair wind power and bitcoin mining, now being used for AI training workloads. 

​

Lastly, she spoke about hardware lifecycles. There is significant e-waste that is generated from AI. There are constantly new models being released, and reusing old hardware becomes limited. She points out that GPUs typically have shorter functional lifespans and often require more updates than CPUs.