How To Build a Large Switch
A practical guide
To make this project interesting, let’s build a switch to meet the following requirements:
- 150 – 500 Optical Inputs
- 150 – 500 Optical Outputs
- Each input/output handling 10Gb/s – 40Gb/s traffic
- Non-blocking
- Remotely reconfigurable.
If you look at building a switch in this class, one of the first conclusions you come to is that the solution has to be an optical switch. While electrical switching performance is constantly improving, the design of an electrical switching fabric of this size and speed would have to include the specification of a major power substation next door to keep it fed. The transparency of optical switching to data rates is a major advantage when trying to handle such a large number of 10Gb/s to 40Gb/s streams.
Restricting ourselves to optical switches still leaves a large number of options. Over the last 10 years, a large number of optical technologies have been proposed for optical switching. Many of these have actually been built, either as research projects or products. Most of them make use of some non-linear material property to build a simple 1 x 2 switch. You can view the 1 x 2 as the transistor of their optical switching world. Using the mathematical magic of powers of 2, these can be cascaded to make 1 x n switches where n is large. Hopefully, the 1 x 2 switches will be very low loss, or the 1 x n cascades will be unusable.
Thus, a number of technologies are capable of building 1 x 150 to 1 x 500 optical switches. A few of these would be sufficiently small and low loss to be considered for the next step, our large n x n switch. Here is where we confront a fundamental problem which was apparent to the electrical ASIC developers years ago. It’s not the logic elements, it’s the interconnects that kill you. As an example, let’s construct a 150 x 150 switch using 1 x 150 (and 150 x 1) parts. We will need a 1 x 150 for each of the inputs and a 150 x 1 for each of the outputs, or 300 devices. In addition, we will need 150 times 150 or 22,500 fibers to connect all of these devices together!
A 500 x 500 switch, built in the same manner, would require 250,000 interconnects. An alternative approach would be to construct small non-blocking m x m switches and using a three-stage network to build the larger fabric. However, here too the loss in a three switch cascade combined with the large number of interconnects is a major problem when trying to build a practical switch.
Fortunately, we are working in the optical, not electrical, domain. Unlike wires, in free-space beams of light do not short out when they cross. In optics, there is a technology which can remove the interconnects as a problem, MEMS. MEMS leverages the simplicity of free-space optics using lenses and mirrors to relay light from the face of an input fiber to the face of an output fiber with an absolute minimal loss of light. The lenses and mirrors have millimeter scale dimensions with micrometer scale features. For example, optical micro-lens might be less than 1 millimeter in diameter. Manufacturing of these devices benefits from the scalability and uniformity of semiconductor silicon batch processing, enabling low cost fabrication and high performance operation.
A simple 1 x 2 MEMS switch could consist of a single pop-up mirror which would redirect the light beam from one output to another. This has been extended to the classic crosspoint switch architecture with n input rows and n output columns with a pop-up mirror diagonal across each intersection. Since the mirrors are so small, a modest n x n switch can be built on a single chip with no interconnects required. However, this 2D approach has its limitations. Mathematics is working against us. The number of mirrors required is the square of the number of inputs or outputs. Thus, it requires 1024 mirrors for a 32 x 32 and 250,000 for a 500 x 500 switch. Even at only 1 millimeter square, that’s a very large piece of silicon. As the switch size increases, there are also major issues of loss and reliability. This is not a practical approach.
If we get out of the plane and into 3D MEMS, these problems are solvable. The 3D MEMS architecture relys on much more precise and flexible control of the mirror positioning to redirect the input light beam to one of many outputs, not just two. Thus, each mirror could become the equivalent of the 1 x 150 or 1 x 500 switch mentioned above. Utilizing the same switch architecture described above, a 500 x 500 switch could be constructed with only 1,000 mirrors instead of 250,000. Furthermore, all of the 500 input switches (mirrors) could be fabricated on one piece of silicon not much bigger than a postage stamp.
3D MEMS is the only technology capable of building large, high reliability, low-loss, non-blocking switches. Unfortunately, while it is easy for me to say that we just use mirrors to bounce the light beam from input to output, it is very difficult to do well. It has taken Glimmerglass six years and tens of millions of dollars to perfect the technology and create the highly-praised products we now ship. This is not a do-it-yourself project.
