Faculty Seminar: A Data Center Energy Solution and the Science of Decision-Making

Anthony Rizzo (l) and Bryce Ferguson, both assistant professors of engineering at Dartmouth.

Data-Center Energy Crisis? Let There be Light

Rizzo, principal investigator of the Rizzo Integrated Photonic Systems Laboratory at Dartmouth, launched the seminar with his presentation, "Addressing the Data Center Energy Crisis with Light."

As Rizzo explained, data centers are on a dangerous collision course. While the compute demands of artificial intelligence are growing at a rate Rizzo calls "insanely fast," the transistors that power our computers have gotten about as small as is physically possible.

In other words, the 60-year reign of Moore's Law—which says the number of transistors in an integrated circuit doubles about every two years—has come to an end. Future improvements to conventional circuitry are unlikely to meet AI's burgeoning demand for raw computing power.

Meeting this challenge is further complicated by the nature of today's data centers, Rizzo said. The huge new data centers designed for AI workloads contain literally thousands of computer racks, all filled with servers containing multiple GPUs and CPUs. These servers—plus the communications networks that let them share data—have a thirst for electric power so deep, it's straining local power grids.

Those data-center networks are also the source of serious bottlenecks. Even if engineers could design smaller transistors, Rizzo said, these bottlenecks would still prevent AI workloads from scaling as needed.

The solution? Silicon photonics.

As Rizzo explained, these are specialized devices created from the same silicon used for conventional processors. But instead of working with electrical signals, silicon photonic devices convert those signals into light on chips, send the information over optical fiber, and then use devices known as Germanium photodetectors to reverse the process, turning light back into electricity.

All this can be done at chip scale with extreme efficiency, Rizzo said. For one, a photonic device capable of encoding information into 32 colors of light requires only a single laser. And while the chip itself is tiny, measuring just 100 micrometers across, it can shift up to 500 billion bits of data per second.

Introducing photonic devices into data centers, Rizzo said, will make their systems "distance agnostic." In essence, a big data center will perform like a single computer, with distance no longer affecting either bandwidth or energy consumption.

"These ultrahigh-bandwidth and energy-efficient photonic interconnects are going to revolutionize future computing systems," Rizzo said. "Doing this ultradense heterogenous integration of electronics with photonics is absolutely critical to getting the best possible system performance."

Pricing or Persuasion? Try Both

Ferguson, who directs Dartmouth's Multi-Agent Decision, Control, and Autonomy Theory (MADCAT) Lab, followed with a presentation entitled "Algorithmic Incentive Design via Pricing and Persuasion."

A motivation of Ferguson's research is a concept known as "the tragedy of the commons." It posits that smart local decisions made by individuals can lead to bad results at the global level.

One example is the situation of the Grand Banks. This underwater shelf, located off the coast of Newfoundland, Canada, is a natural home to cod. During the early and mid-twentieth century, local fishermen sought to catch as many cod as possible. At the individual level, it made sense. The more cod a fisherman caught, the more money they could earn. But at the global level, it led to literally decades of overfishing. By the early 1990s, the Grand Banks' cod population had been reduced to near extinction. What had been good for fishermen individually became a disaster for both the cod and the collective livelihoods of local fishermen.

To solve this kind of dilemma, Ferguson is working with approaches known as algorithm game theory. As the name implies, the approach combines game theory, which models individual behavior and then predicts collective results, with algorithm design, which creates decision-making rules to get desirable results.

Another foundation of Ferguson's work is known as the price of anarchy. This formal metric offers a ratio: the quality of strategic output divided by the quality of optimal output. For example, imagine the average time it takes a driver to commute to work when they choose their own route. Now imagine that the trip, if it had been optimally scheduled, would have taken less time on average. The difference between the two times is an example of the price of anarchy.

To overcome this difference, proponents have tried three approaches: policy, or setting and enforcing rules; incentives, or rewarding certain behaviors; and persuasion, which aims to alter information and beliefs. However, none have really worked.

To drive better results, Ferguson and his colleagues have devised a solution that combines information-revealing mechanisms with monetary incentives. In other words, they reveal information to shape beliefs while also charging them for certain actions.

Real-world examples already exist. Think of the way Uber and Lyft charge more for a shorter wait. That combines information—how quickly a car is available—with a monetary incentive—it's cheaper to wait.

Ferguson said his solution also overcomes another challenge. Prior research has found that revealing information can actually make performance worse.

"On their own, incentives can be harmful, and just revealing information can be harmful, too," Ferguson said. "But when you design the two together, you can reduce the possibility of harmfulness."