As the computing industry has grown dramatically in the past few decades, so has the environmental footprint created by such technologies as artificial intelligence and cloud computing. The industry is now looking beyond the operational carbon footprint, or the carbon created by a computer system. Instead, it is considering the life cycle carbon footprint, which includes both the operational and the embodied carbon footprint. The embodied carbon footprint is the carbon created in the manufacturing of a smartphone, an AI system, or other hardware. Before you can optimize computing’s carbon footprint, though, you must be able to measure it. And measuring the embodied carbon footprint has been challenging to date.
Measuring the embodied carbon footprint
To solve this dilemma, researchers from Meta and Harvard University have created an architectural carbon modeling tool (ACT) to quantify and optimize the embodied carbon footprint. Traditionally, system architects would design a given computer system to optimize its performance, power, and chip area. With ACT, system architects can design and optimize hardware to minimize carbon yields as well as to optimize for performance and energy efficiency. For example, some smartphones contain hardware components that aren’t needed for all user applications. These extra components add to the carbon footprint of that phone. So when optimizing for application use cases, a system designer can use ACT to predict the additional carbon cost of a particular functionality. In doing so, the system designer could incorporate only those hardware components that are critical for the desired applications, consequently optimizing the life cycle carbon footprint of the phone. Put differently, system designers can now quantify the trade-offs between given features and their contribution to the phone’s overall carbon footprint. And more broadly speaking, the community can use ACT to measure and ultimately reduce the environmental footprint of technologies such as AI, as system architects can design greener AI systems from the start.
A first-of-its-kind tool, ACT is open sourced so that system architects and application developers have the necessary model and metrics to quantify the carbon footprint of their solutions too. Using the three tenets of sustainable design — reduce, reuse, recycle — the researchers have demonstrated how future computer systems can achieve strong performance and efficiency in an environmentally sustainable manner.
“Our vision is to be able to put the equivalent of a nutrition label on any computing equipment to show how green it is,” says researcher Carole-Jean Wu.
The best way to reduce the carbon footprint, according to the researchers, is to eliminate carbon emissions during the system design and production phase. One way to do so is to design leaner hardware accelerators to balance performance, power, energy, and carbon emissions. The team looked at the mobile AI inference as a use case for how to reduce emissions and explored the trade-off between performance, power, area, and carbon. They found that while AI accelerators in newer process nodes achieve higher efficiency, they also incur higher embodied emissions.
Therefore, to curb and reduce the growing footprint from hardware advancement, system architects must design systems under a strict carbon budget. By considering the carbon footprint of different process technologies, including compute logics, memory, and storage, system architects can minimize the overall carbon footprint of a computer system at the design time.
The researchers also investigated how, from a sustainability perspective, it is prudent to reuse products. The optimal hardware design, between programmable and reusable versus specialized hardware, varies based on the availability of renewable energy during manufacturing. Depending on the hardware manufacturing locations, embodied carbon varies.
One way to design sustainable platforms more broadly, for example, is for mobile vendors to offer different chipsets to users based on the fabrication characteristics and availability of renewable energy in the field. Simply providing carbon measures for system hardware components is already an effective way to enable sustainable computer design because system designers can consider carbon as an additional dimension in the system design space.
For the third tenet, recycle, the researchers proposed designing modular systems and extending hardware lifetimes. Generally, extending hardware lifetimes can minimize embodied emissions as fewer systems are produced over time. With modular systems, for example, it’s possible to update or upgrade the memory unit without decommissioning the whole system.
With the ACT framework, the researchers hope to equip the industry with a tool that can help reduce computing’s overall carbon footprint, including carbon emissions produced by AI, one of the key application domains. “With our ACT tool,” notes Wu, “we are able to design low-carbon systems while meeting the optimization objectives of high performance and high energy efficiency, and develop computer systems in an environmentally sustainable way.”