Fostering energy efficiency in search systems: an interview with Ryu Dong-ho

The growing usage of technology in our daily lives has inevitably led to its greater environmental footprint. In response, search engineering companies have made an effort to make search systems more energy-efficient as part of their broader ESG commitments. One such company is AXZ, a spin-off company of Kakao that operates the major Korean search service Daum. In this interview, Ryu Dong-ho at AXZ's AI Search Service Development Team discussed how his team works to increase the energy efficiency of search systems and his personal insights on the role searching systems play as an indirect contributor to climate awareness.

I’ve heard that power consumption in data centers is increasing due to large-scale search services and AI models. What efforts are being made to improve the search system?

While providing accurate search results to users is important, delivering those results quickly is also crucial. That’s why we devote a great deal of attention and effort to operating the system efficiently. As a result, operating a portal service like this requires a huge number of servers. However, we don’t just blindly use a lot of hardware; we make efforts to scale up quickly when user traffic demands it, and when it doesn’t, we reduce unnecessary servers to maintain an optimal number that doesn’t disrupt the service. 

My specific role involves search modeling—specifically, work related to search ranking. We use machine learning to create models that use data to train computer models and then rank search results based on those models. Recently, very large and powerful models called LLMs have emerged, and while it would be great to use them, if we were to have the LLM respond to every single query or request that users make directly in the service, it would take a long time and inevitably consume a huge amount of power.

So, we’re working on a process where we generate data, use the LLM to create and experiment with it, and then, before deploying the service, we retrain the results of that training on a smaller-scale model. This allows the service to achieve similar results even with a smaller model, so I think we can save some power in that regard.

Are there any efforts within your team to measure or manage the carbon footprint generated during the actual process of training and operating these models?

We do measure our carbon footprint, but these days, ESG management and related aspects are essential for running a business, so we address them at the company level. Since optimizing these services helps reduce costs for the company, we’re not just measuring the carbon footprint—we’re actively striving to use resources efficiently.

What role do you think the data and algorithms handled by axz play in better conveying information related to eco-friendliness and carbon reduction?

Since the service we’re developing is a portal where people can come together to share opinions, we’re working to create a platform where everyone can gather to freely discuss and exchange ideas. So, while we’re not focusing exclusively on carbon footprints or environmental issues, we’re striving to create a service where people can freely communicate about broader societal issues and problems—including those aspects—and view the results objectively, without bias.

As a search service developer, what are your thoughts on the idea that technology can contribute to solving the climate crisis? What are the realistic possibilities and limitations?

I don’t think a search service can have a direct impact on its own, but I believe that while science and technology have certainly created problems in addressing climate and environmental issues, they are also the means by which we find solutions. And while search itself may not contribute directly, I think it can serve as a gateway for people interested in these issues to explore them. By providing more accurate, diverse, and high-quality search results to those with such curiosity, I believe we can make an indirect contribution.

As data center energy efficiency improves and the use of eco-friendly power increases, could the way we develop these services change as well?

This applies not only to our service operations but also to development. When we develop, we need equipment called GPUs, but these GPUs are actually very hard to come by these days, so we don’t have the luxury of using them freely. That’s why the company now provides an environment where multiple people can develop and experiment using GPUs in the form of a cloud service.

When we develop there, we don’t just use the resources carelessly; we take only the resources we absolutely need to conduct our experiments, and once our work is done, we return them so that others can use them—we’re making an effort to operate in that manner.

Among the projects you’re currently working on, are there any cases where you felt this technology brought about socially meaningful change?

Yes, although this might not be directly related to the environment, people use portals to read news articles and share their opinions through comments. During election seasons or similar periods, user activity tends to increase significantly. So, I believe the platform serves as a channel for users to exchange opinions, take an interest in social issues, and communicate with one another on those topics.

Finally, in a world where climate crises and carbon neutrality are becoming increasingly important, do you have any future environmental goals related to the work you’re doing?

Environmental issues are a crucial and sensitive part of our daily lives, so I believe everyone should have some interest in them, even if they aren’t directly involved in that field. These days, the automotive industry is making efforts to introduce electric vehicles and hydrogen-powered cars. Additionally, the use of machine learning—which is what I specialize in—is increasingly being applied to areas like autonomous driving, leading to a significant expansion of software-related aspects in the automotive sector.

Ultimately, the scope of work involving machine learning and AI—training models on data and applying them to services—is steadily expanding. Since advancements in one area can influence and drive progress in others, I believe that if I do my job well, it will ultimately contribute to those areas as well.