Silicon photonics -- the long-term copper replacement
In theory, silicon photonics could solve some major problems associated with the continued use of copper interconnects. One of the biggest problems with copper wire is that it doesn't scale nearly as well as other vital parts of a modern CPU. Past a certain point, it becomes physically impossible to make copper wires any smaller without compromising their performance and/or lifespan. In theory, optical interconnects could transmit data for far less power while simultaneously moving information much more quickly.
Silicon, unfortunately, is a poor native medium for optical devices. Because silicon lacks a bandgap and the scales of manufacturing are so different (optical waveguides and other components are far larger than the silicon CMOS devices they interact with), designing solutions that could scale effectively and affordably, integrate into existing CMOS manufacturing, and rely on silicon rather than costly alternative materials like gallium arsenide has proven extremely difficult.
The reason so many companies have pushed to bring this technology to market, despite the slow pace of progress, is that silicon photonics is generally believed to be necessary for exascale-level computing. Right now, copper and fiber typically split the transmission market by distance. Short-run cables between servers or racks tend to use copper, while longer distances rely on fiber.
The chart above is from an Intel presentation on silicon photonics, but it illustrates the cost and power consumption targets that manufacturers are trying to hit. The long-term roadmaps for silicon photonics enables bandwidths and energy/bit of information ratios that no copper signaling can match. Bringing power down from 75 picojoules to 250 femtojoules is a reduction of multiple orders of magnitude.
As part of that effort, IBM's research teams have worked to reduce the process node that it used for circuit design. This slide from 2012 shows how 90nm - 65nm represents the "sweet spot" for these kinds of circuits. While we're used to smaller nodes offering substantial benefits to traditional CPU transistors, other kinds of components don't see the same benefits from scaling to smaller process geometries. IBM's documentation refers to sub-100nm manufacturing, implying that the company standardized at 90nm or 65nm.
IBM isn't giving timelines for when we might see more devices shipping with on-chip silicon photonics, but we can predict how the technology will roll out. Current cutting-edge designs put the optical components on the same physical package as the CPU, or at the edge of a motherboard. This makes the hardware useful for server-to-server linkages or possibly for peripheral connection. We expect to see silicon photonics roll out first in the HPC and scientific computing industries, where the sheer scale of many build-outs makes the power conservation critical and government grants are available to ease the cost of initial deployment. After decades of work, silicon photonics might seem like just another pie-in-the-sky idea that sounds great on paper and never pans out -- but from HP to Intel to IBM, progress is happening in this field. Hardware may not roll out today or next year, but optical signaling is going to play a part of computing's future -- in the datacenter, even if nowhere else.Tagged In
Data Signaling Exascale Cpu Intel Silicon PhotonicsMore from Extreme
