For non-scientists this webpage may raise questions concerning the topics of JOLLYBEE. In a try to answer them without scientific and technical terminology, the following sections address the most important questions that may have come to the mind of the interested reader.
What is neuromorphic computing?
…and why do we care?
The Internet Explosion has led to a massive volume of 1.1 Zettabyte inundating the global communication infrastructure as early as 2017. This volume shows a strong tendency to increase further. More importantly, this volume goes along with a skyrocketing energy footprint for the ICT sector, equivalent to an energy consumption that is expected to reach more than 8000 TWh in 2030. This energy footprint, that applies to the transport of data, its processing and its storage, corresponds to a significant share of up to 21% in global energy consumption by the end of the decade – if no further measures are taken to reverse the current trend.
However, an interesting fact can be observed when it comes to data processing: While today’s supercomputers offer a processing power of more than 500.000 TFlop/s at a power consumption of 30 MW, the human brain facilitates ~2000 TFlops/s at only 0.00002 MW. This fact is best underpinned by the legendary Go-Match between Google’s AlphaGo and Go-master Lee Sedol: a machine of more than 1200 CPUs and 170 GPUs (operated at 1 MW) battling against a single human brain (operated at only 20W). This imbalance in computational output per power is caused by the energy brick-wall that current microprocessors are facing.
Thus, there must be a more efficient approach for data processing and computation. This is to be found in neuromorphic computing – a discipline that mimics the information structures of the human brain through artificial neural networks. This novel technology promises an order-of-magnitude improvement in computational efficiency, striving for the ultimate efficiency of a biological neuron.
While the human brain builds on a vast number of ~1011 biological neurons to perform nearly 100 quadrillion (1017) operations per Watt, or alternatively expressed as a sub-Attojoule energy per operation, artificial neural networks will be constructed of a much smaller number of neurons. However, even an implementation with ~25 neurons can already address challenging tasks such as ultra-fast pattern recognition or predictive control at few-nanosecond convergence time.
We would like to point to the following two publications for further reading on the topic by the interested reader: N. Jones, Nature 561, 163 (2018); J. Kendall, Appl. Phys. Rev. 7, 011305 (2020).
Why photonic signal processing?
Key enabling technologies such as photonics, which exploits the properties of light to encode information, and optical signal processing methods to realize artificial neurons, contribute with their unique ability to handle GHz information rates at low energy consumption. This is what JOLLYBEE aims to accomplish.
Although optical signal processing is a rather new field, optical telecommuncations have laid the inroads for this challenging task during the past decades. As a matter of fact, optical ICT technology is omnipresent and addresses medium-and short reach domains such as inter-city metro links or optical access networks that reach till the households. Furthermore, datacenters, which consitute the backbone of the “cloud”, are employing a massive number of optical links in order to interconnect their servers, their circuit boards and even processor cores. Even mobile networks would not function without optical telecommunications: As soon as the radio wave reaches the antenna site, it is digitized and transported over fiber optics to a centralized processing unit that serves the signal reconstruction and reception. As such, more than 90% of the bits that transmitted over ICT networks are passing over optical links to date.
This little clip shows the operation of an optic link, such as it would apply in a photonic synaptic interconnect. Here, a water jet composes the fiber channel. Light, in this case at a red wavelength, is guided along the jet as it bends down to the reservoir with a receiving photodiode. In this way classical music can be transmitted over a water jet – as long the transmission isn’t interrupted.
Why optics in particular? What are other advantages?
…simply because optics provide a better performance than electronics. An important fact is that the particles of light, the so-called photons, are much more “robust” for communications than their electrical counterparts, the electrons. It is possible to transmit ~10.000-times more data over a single optical fibre compared to an electrical cable. At the moment, it is possible to transmit up to 10 Pb/s over an optical waveguide, for example an optical fibre that is as thin as a human hair. At the same time, it allows us to operate networks over much longer distances, e.g. more than 10.000 kilometers, and with much higher energy efficiency.
The latter becomes especially important as the energy footprint of ICT becomes larger and larger. Every transmitted bit over metro and core networks generates ~100 bits within the datacenters. Content generation and processing is surging at immense rates, which already exceed these found for transmission capacities. The handling of data streams within the datacenters, which mark the hotspots of out modern ICT infrastructure, is to be best perfomed optically at lowest possible energy consumptions in the order of 5 pJ per transmitted bit.
A similar challenge applies to artificial neural networks that build on neuromorphic photonics: high information densities to be processed with inter-networked optical neurons at minimal energy consumption.
The image besides shows an optical transceiver under the microscope, as it would constitute a synaptic receptor or emitter. It accepts local electrical data to transmit it to the outside world via a thin optical fiber. At the same time, it receives photons, in this case at a red wavelength that carries a high-speed information signal.