Super-fast optical interconnects

Universal chip-to-chip optical interconnects are delivering off-package communication at the bandwidth density and energy cost of in-package electrical incumbents.

BY VLADIMIR STOJANOVIC FROM AYAR LABS

Data dashes through today’s optical fibre system networks before it slows to a crawl at copper interconnects. These bottlenecks occurs at copper pins and wires on circuit boards, where electrons transmit data at far lower speeds. So great are these delays that there will come a time when copper interconnects will have to be replaced by optical signalling.

Offering a universal solution to this issue is our team from Ayar Labs of Emeryville, CA. Founded in 2015 as a spin-out of three universities  Massachusetts Institute of Technology; University of California, Berkeley; and University of Colorado, Boulder  we are renowned for our pioneering work in micro-ring chip architectures.

Figure 1. Large penalties will be incurred for leaving the chip, package, and board (source: Gordon Keeler, DARPA MTO, ERI Summit 2019).

We have developed a chip-to-chip optical input/output (I/O) interconnect technology that addresses several weakness associated with electrical interconnects and their scaling limitations. We tackle weaknesses associated with signalling speeds and pin count, which both apply the brakes to electrical I/O connections, and power consumption that is increasing at an unsustainable rate.

The latter concern should not be taken lightly, as it will not be long before the power drawn by the off-chip I/O will account for almost all the power consumption of the package. When this occurs, it will be infeasible to use of electrical I/O interconnects, which are primarily made of copper. By then, there will need to be have been a shift to chip-to-chip communications based on photonics, a technology that will eliminate electrical I/O bottlenecks.

Introducing new photonic I/O architectures will also satisfy growing throughput demands. There is an emergence of heterogeneous computing, involving central processing units (CPUs), graphic processing units (GPUs), field-programmable gate arrays (FPGAs), neural network accelerators, and resource pooling on the memory side. This trend demands more I/O at the application level, while electrical I/O is running into pin count, signalling and power limitations.

Providing an impetus to act sooner rather than later, the penalties for leaving the chip, package, and board are on the up (Figure 1). This begs the question: will the 112-Gbit/s serializer/deserializer be the last long-range electrical I/O solution? In field deployments, system integrators are already seeing the limitations of 112-Gbit/s long-range electrical connections  they are incapable of spanning the signalling lengths required for off-board, rack-scale communication.

To evaluate the performance of the I/O technologies, we use a figure of merit involving quite a few terms. It considers the bandwidth density from the edge of the I/O solution, multiplied by its associated energy efficiency. This provides a yardstick that captures the Gbit/s per millimetre per pJ/bit, as a function of distance. When crossing the boundaries from package to board with I/O technologies, this metric plummets – there is a gap of about four orders of magnitude between in-package interconnect solutions and state-of-the-art optical solutions that provide off-board connectivity.

Into this challenging space comes our new optical I/O technology. It delivers a universal I/O solution that provides off-package communication at the bandwidth density and energy cost of in-package electrical incumbents.

Figure 2. Ayar Labs’ in-package optical I/O system architecture (source: Ayar Labs).

An in-package solution
To pull off an optical I/O, certain requirements must be fulfilled (see Figure 2). We meet them with our own optical I/O system architecture. This is a chiplet-based solution that is co-packaged directly with the host system-on-chip (SOC) within a multichip module (MCM) package. By incorporating a monolithic electronic-photonic CMOS chiplet, which we call TeraPHY, we realisea flexible electrical I/O interface adapted to the host SOC, whether the geometry is wide parallel or high-speed serial. Adopting this approach offers flexibility, giving a choice between a silicon interposer and an organic substrate for the package.

One of the merits of our single-chip solution is that it enters a manufacturing ecosystem already established for MCMs. With our architecture, we keep the laser supply outside the module. This simplifies packaging, improves laser reliability, and alleviates issues related to the operational temperature mismatch of SOC packages and laser optical supplies. Due to these advantages, we realise further integration of lasers into a multi-port, multi-wavelength laser module solution, which we refer to as SuperNova.