On-Chip Filters Could Drop the Power Draw of Future Photonic Chips by a Factor of a Million
Swapping out heating elements for a "near-zero power consumption" tuning alternative could be a real winner for high-performance compute.
Researchers from Oregon State University and Baylor University have come up with an approach which could cut the power required by next-generation photonic computing chips considerably — by a factor of a million, in fact.
"Alan is an expert in photonic materials and devices and my area of expertise is atomic layer deposition and electronic devices," explains John Conley, referring to corresponding author Alan Wang, of what went into the team's work. "We were able to make working prototypes that show temperature can be controlled via gate voltage, which means using virtually no electric current."
Temperature control of photonic chips is, interestingly, the most energy-inefficient part of the design. Frequently used for high-speed communication and considered one of the front-running technologies to take over from traditional electronics in high-performance computing systems, photonic circuits only work when their operating temperature is kept stable — which usually involves more energy into heating the chips than into actually generating or processing light pulses.
While, admittedly, we're talking milliwatts rather than megawatts, it's still a considerable burden. "[It] might not sound like much considering that a typical LED lightbulb uses 6 to 10 watts,” Wang admits. "However, multiply those several milliwatts by millions of devices and they add up quickly, so that approach faces challenges as systems scale up and become bigger and more powerful."
The team's approach to tuning the wavelength of light in a photonic chip ditches wasteful thermal heaters in favor of an on-chip gate-tuning system which offers a large wavelength coverage and tuning across the 1543-1548nm range with no gaps — and, more importantly, near-zero power consumption estimated at around one-millionth the draw of a heater-based alternative.
“Our method is much more acceptable for the planet,” Conley says in comparison with traditional approaches. "It will one day allow data centers to keep getting faster and more powerful while using less energy so that we can access ever more powerful applications driven by machine learning, such as ChatGPT, without feeling guilty."
The team's work has been published in the journal Scientific Reports under open-access terms.