Chip Interconnects are Going Optical
Though copper provides high-speed connectivity, it has limitations. One of the main constraining factors is distance, but even delivering higher bandwidths has its limitations. The move to optical connections is the holy grail since distance becomes less of an issue and there are advantages when it comes to issues like crosstalk.
Unfortunately, in the past, fiber was also more costly and difficult to deploy. However, this is rapidly changing as we move from external pluggable connections to on-chip fiber connectivity (Fig. 1).
One main reason for pushing the connectivity boundaries to fiber is that large-scale, artificial-intelligence (AI) acceleration requires lots of compute power, a huge amount of storage, and a way to connect it all together. These systems don’t fit in one box no matter how much we shrink things.
The data centers of yore, with mainframes and raised floors so that cable and cooling could fit underneath, have been dwarfed by massive data centers needing fiber optics to connect multiple racks containing the compute, communication, and storage hardware.
Moving Past One Wavelength Per Fiber
Typically, fiber connections used a single wavelength, and speeds were fast. However, fiber can also support multiple wavelengths. The challenge in accomplishing that is twofold: First, there must be a way to multiplex the signals, and second, the support across the wavelengths involved has to be consistent.
On this front, Lightmatter now delivers 16 bidirectional, dense wavelength division multiplexing (DWDM) optical links on one single-mode fiber (Fig. 2). It delivers a bandwidth of 400 Gb/s in each direction. A “closed-loop digital stabilization system actively compensates for thermal drift, ensuring continuous, low-error transmission over wide temperature fluctuations.”
The more impressive aspect of Lightmatter’s Passage announcement is that it’s designed to operate at the chip level. In the past, the fiber connections were outside of the chip that had a copper connection to the optical transceivers. The company’s approach is inherently polarization-insensitive. This is an issue when one must consider connections and mechanical stress.
At the chip level, a chiplet approach is taken with the M-series, whereby the optical support is put on the chiplet with direct fiber connections (Fig. 3). With the L-series, an interposer provides connections to the optical transceivers on the periphery using fiber.
Single Fiber with 16 DWDM Optical Links
Shrinking the Optical Modulators
Mach-Zehnder modulators (MZMs), electro-absorption modulators (EAMs) and micro-ring modulators (MRMs) are used in optical systems now (Fig. 4). Lightmatter leverages MRMs. These are more compact and allow the transceivers to fit on-chip. The other options are larger and require implementation outside of the chip.
Shrinking the size of the modulators is just part of the puzzle. Reducing power consumption is key as well. Luckily, power requirements shrink significantly.