Optical Computing 101: How Does It Work?

Article By : Sally Ward-Foxton

With optical devices hitting the headlines for efficient AI accelerators, we ask: how does optical computing work?

How does optical computing work? This technology has been gathering momentum with two startup companies (Lightmatter and Lightelligence) demonstrating optical compute chips designed for AI acceleration, and handful of others still working on it. Lightmatter, for one, last week unveiled at Hot Chips 2020 its AI photonic processor that uses light to compute and transport data. Optical compute uses barely any power and operates at the speed of light (in silicon), promising big improvements over transistor-based compute for workloads like AI inference. The enabling technology for optical compute is silicon photonics, the field dedicated to sending infrared light through silicon structures on a chip. Driven by optical communications applications, particularly for data center infrastructure, silicon photonics is used to integrate optical components onto a silicon chip to take advantage of CMOS’s low cost and scalability and CMOS devices’ ease of fabrication and assembly.
Recommended: Optical Processor Targets to Accelerate Compute for Next-Gen AI
While the concept of optical compute has been around for a while, it’s only the development of silicon photonics that has really made it possible in the last couple of years. Previous attempts to create an optical version of the transistor using conventional optics were unsuccessful; the work of companies like Lightmatter and Lightelligence does not try to force light to work like electrons in a transistor, instead using the fundamental properties of light to perform MAC operations. At the most basic level, an optical computer encodes data as the brightness of light. “Imagine you have a flashlight with some transparent material in front of it, and you can control how much light makes it through – that’s a modulator,” explains Lightmatter CEO Nick Harris. “We use modulators to control the brightness of light, and we can encode information that way. What that ends up doing is a multiplication. You’re multiplying the signal by the amount of transparency.” Addition, at a basic level, is simply a matter of joining two waveguides together (waveguides are the “wires” light travels down – in Lightmatter’s chip the waveguides are around 300 by 200 nm). These concepts combined can make optical multiply-accumulators (MACs), and used to create larger MAC arrays that are required for multiplying matrices together, the key operation required for computing today’s AI inference workloads. Modulation of light Of course, devices on a real chip are somewhat more complicated. Lightmatter’s optical compute architecture uses Mach Zehnder Interferometers (MZIs) as modulators. MZIs take coherent light (a beam of light with all the light at the same wavelength and phase), split the beam of light in two at a Y-junction, and then apply different phase shifts to each half-beam. The half-beams are then recombined at another Y-junction. The different phases causes constructive or destructive interference which effectively modulates the light by the amount desired.
How does optical compute work? MZI diagram
Mach Zehnder Interferometers use the phase shifts and the principles of interference to modulate light (Image: Lightmatter)
How is the phase shifted created in silicon photonics? Well, explains Lightmatter VP Engineering Carl Ramey, there are several different ways. “Probably the most common way is through a heater,” Ramey said. “By applying heat to the wave guide, you can actually change the index of refraction and you can affect the speed of light through the wave guide. This in turn causes a phase shift that’s proportional to the change in temperature and also proportional to the length of the segment that you’re heating.” Unfortunately thermal phase shifters are in general too slow for high speed computing due to the thermal time constant of the segment being heated (they can operate in the kHz range, Ramey said). Another photonic phase-shift architecture adds dopants to the waveguide to form a p-n junction that can be actuated at high speed, but this type of device is too large for a compute array, Ramey said, and they suffers from high losses. The limitations of other techniques for building silicon phase shifters led Lightmatter to take its cue from MEMS and use a mechanical device. Actually, a nano-optical electromechanical system, which is given the acronym NOEMS. In a NOEMS phase shifter, the waveguide is built like a MEMS structure, with the material above and below etched away. Applying a charge to panels above and below it can mechanically bend the waveguide by a small amount, enough to affect its refractive index and shift the phase of the optical signal passing through. Ramey said Lightmatter chose NOEMS because they are extremely low loss, the static power dissipation is close to zero, the dynamic power is small and they can be actuated at high frequency (hundreds of MHz). “If you can modulate the phase, you’ve created a multiplier that runs at the speed of light and the only energy consumed is that of the phase shifter and the tiny amount of optical loss through the wave guide itself,” Ramey said. “This is a powerful concept. The calculation is occurring with no fundamental energy required. As fast as the operands can be fed into the input, results are available at the output with nearly zero energy consumption. And this is independent of process scaling or voltage.” Multiplexing signals A unique property of optical computing is the ability to multiplex. This idea is already used in fiber optic communications, where using different wavelengths of light for each signal means multiple signals can be sent down the same waveguide at the same time. The same theory applies to optical computing: multiple AI inferences could be performed simultaneously by encoding each set of input data in different wavelengths of light. While in theory more than a thousand channels are possible, in practice this is limited to eight inferences at the same time, due to the current limitations of laser technology. Doing multiple inferences at the same time effectively makes the optical chip perform the work of eight chips at once and is a powerful way to increase compute performance.

Leave a comment