Here's one way past Moore's law: Chips that mix photonics and electronics
Boffins see promise in optoelectronic accelerators running faster with less energy than traditional components
As Intel, TSMC, and Samsung race to prove that Moore's law is still relevant with faster and more efficient chips, researchers in China have pointed to a growing body of research revealing one way silicon slingers can achieve higher levels of performance: matrix math accelerators that integrate electronic circuits and photonics.
This conclusion about the promise of so-called optoelectronic chips was published this month in the peer-reviewed Advanced Photonics journal. The review was conducted by two academics at the Chinese Academy of Sciences and Peking University in China.
The eggheads, Pengfei Xu and Zhiping Zhou, argue that optoelectronic chips may not make sense for all types of general-purpose computations.
But when it comes to performing matrix operations – which are crucial for high-performance workloads such as AI, simulations, and graphics rendering – optoelectronic chips "have the potential to surpass the computation performance" of traditional electronic processors "in terms of energy efficiency, computational power, latency, and maintainability," according to the duo.
"We believe that silicon-based optoelectronics is a promising and comprehensive platform for general-purpose matrix computation in the post-Moore's law era," Xu and Zhou wrote.
Don't stop us if you've heard this before, because we have, too, as recently as June, back in 2015, and in between and beyond. Experimental chips driven by light routinely crop up, it seems. Xu and Zhou's paper is an analysis of where we're at after so much promise from the industry, and specifically "the recent progress in photonic matrix computation."
- IBM wants to simplify 3D stacked chip manufacturing
- First details on TSMC's 2nm node: Chipmaker reveals nanosheet transistors
- Take this $15m and make us some ultra-energy-efficient superconductor chips, scientists told
The boffins believe optoelectronic chips have the potential to exceed regular computer chips in three key ways: data transfer speeds, lower latency, and lower energy consumption. That is to say, it's not just matrix operations light can help with; there's fast communications, too.
"Instead of power-hungry electronic transceiver circuits, on-chip optical transceivers are good alternatives for low-energy-budget interconnections and boosting the data movement among the processors, memory, and peripheral hardware," the duo wrote.
The academics point to several signs of progress in developing the technologies necessary to make optoelectronic chips a practical reality.
These include a 2015 demonstration of an integrated photonic-electronic chip with than 70 million transistors and 850 photonic components that had an aggregated memory bandwidth of 55Gbit/sec. There's also a processor, made by Boston startup Lightmatter, consisting of ASICs and a photonic core that can achieve high performance for matrix-vector multiplication operations.
"By exploiting the advantages of light in linear matrix computations, the photonic core is excellent at disruptively improving the computing performance, while the electronic circuits are necessary for performing other non-linear operations, such as driver circuits, arithmetic and logic, data storage, and activation function," the researchers observed.
The pair posit that optoelectronic chips can be particularly beneficial for large-scale matrix computation since the total energy consumption of such chips is "merely proportional" to the side-length of a matrix. This, we're told, is in great contrast to electronic processors, where the energy consumption increases in conjunction with the total number of elements in the matrix.
The findings indicate there may be a point in the future when at least some computers and servers use optoelectronic chips to power demanding applications and may be even better than traditional computer chips if Moore's law truly does meet its end some day.
Provided these opto-processors make it out of the lab. ®