Microprocessor communications have received a boost from the integration of electronics and photonics in silicon — a first step towards low power consumption and efficient computing systems. See Letter p.534
Ever since the demonstration of the first microprocessor, 'smaller, cheaper, faster' has been the motto of the microelectronics evolution that enables ever-more-densely packed circuits to speed up computer performance. But a bottleneck now exists in terms of speed and power consumption for on-chip data communications; for instance, conventional wires lose energy and reduce the communication speed. A full transition to optical-link technology using photons would overcome these limitations, because photonic devices have no speed constraints and use less energy than conventional electronics1. Because silicon is the main material for complementary-metal-oxide-semiconductor (CMOS) technology, which is widely used by the electronics industry, it has been the subject of intense research in photonics, giving rise to what has been called the silicon photonics age2. Using the same material for electronics and photonics in a single circuit could increase performance and reduce the power consumption of integrated chips. On page 534 of this issue, Sun and co-workers3 report a big advance in such efforts: a microprocessor that communicates using light.
Despite the growing interest in silicon photonics and the development of efficient integrated devices (circuits) on silicon-on-insulator (SOI) wafers, only a few complete electronic–photonic circuits have been demonstrated. This is because the silicon substrate for photonics is very different from the standard substrates used in electronics — and even slight changes to CMOS technology can degrade the performance of the transistors used in microchips. As such, developing a process to merge electronics and photonics on a single chip is highly challenging4,5.
The first reported strategy for electronic–photonic integration used the 'front-end' approach6, in which transistors and photonic devices are placed on the same layer of a silicon chip. The chosen method of fabricating such chips was based on a custom CMOS electronic process on a non-standard SOI substrate, and enables high-speed light propagation on the chips.
But even if this integration solution was reliable for producing efficient on-chip transceivers for data input and output, developing a more complex on-chip system using state-of-the-art electronics would require large investments of money and technological-process development. Furthermore, the main proposed integration solutions would involve a multichip approach2 in which the photonic and electronic circuits are fabricated independently using different processes, optimized for each technology, and brought together through a bonding technique. Such multichip integration may be the most economical solution in the short term.
Sun and colleagues' 'zero-change' approach3 challenges this thinking. Based on a commercial SOI CMOS process, it uses existing fabrication steps for electronics, accommodating the photonics without any extra development. This allows all existing electronic designs to be used and combined with photonic components without any additional non-standard processes, which may dramatically increase the efficiency and reliability of the resulting system on a chip (SOC).
The authors report several major advances in the field of microprocessor communication. Their electronic–photonic SOC integrates millions of transistors and hundreds of photonic components to form a microprocessor and memory that communicate with each other using light, at a speed of 2.5 gigabits per second (Fig. 1). The photonic components used to guide, code and detect information are based on a combination of materials that are standard in the electronics industry, including silicon, silicon–germanium (SiGe) and silicon nitride — all of which are implemented in CMOS technology.
The researchers used an external source of light to drive the photonic devices at a wavelength of 1,180 nanometres, with which the light could be confined and channelled efficiently within a waveguide in silicon. To minimize leakage of light from the waveguide into the substrate, the authors selectively and locally remove the substrate under the photonic devices. Each of the two optical links between the microprocessor and the memory includes a compact, silicon micro-ring modulator to code the information at one end, and a SiGe detector driven by both processor and memory at the other end.
Although the photonic circuit may seem simple, Sun et al. optimized it to provide error-free transmission with moderate power consumption. Processors produce heat according to how much they are working, creating large temperature changes over time, which could seriously degrade the performance of optical components. But the authors demonstrate that microprocessor communication is robust in their device under different power conditions (different thermal perturbations), thanks to a feedback loop in the SOC.
Sun and co-workers' result is proof of concept for the development of a complex electronic–photonic SOC. However, challenges remain before their zero-change approach can be used for the commercial production of such circuits. First, the on-chip optical communication rate of 2.5 gigabits per second is relatively slow compared with the rate achievable by state-of-the-art silicon photonics systems. An increase in the bandwidth of both the optical modulators and the detectors in the team's SOC would increase the performance of the memory-to-processor link.
Second, a multiwavelength optical circuit may be needed in the future to resolve the interconnect bottleneck. Moreover, much larger numbers of photonic devices and functionalities, including switches, filters and delay lines with low power consumption, will one day become necessary to address the future requirements of computing systems. Finally, it would also be beneficial to scale this approach up for use in multicore processors.