waveguide data links with SiGe SROs

=suggestion =design =explanation =electrical engineering =radio



Coaxial cable is the best way to send a single electrical signal with minimal losses and noise relative to the cable size.

Attenuation is generally proportional to sqrt(freq) * distance / diameter. Potential data rate is proportional to frequency, so for a given distance, the data capacity of a coaxial cable is proportional to its area. Adding more wires to a cable does not inherently improve its bandwidth. In fact, a coax cable (with MoCA 3.0) has a better data rate per cable area and cable cost than ethernet cable IF you need to move data 100m.

The problem is that for shorter cables, the desired frequency is too high for power amplifiers. 5 GHz is about the limit for affordable hardware, and that used to be lower.

And for longer cables, fiber has lower losses than coaxial cable; the only problem is that it's more expensive for shorter cables.

Waveguides have lower losses than coaxial cable, but their minimum diameter is too large at the frequencies low-cost hardware can handle. If our frequency was unlimited, we could use a 50+ GHz signal in a waveguide the size of Cat5 cable.

How about using a bundle of smaller coaxial cables? That approach was used extensively in the past, but it's expensive and fiber is generally better. Still, that is done today for a few applications: as whitequark noted, Thunderbolt 3 uses micro-coaxial cables.



You might ask: if single-mode optical fiber can transmit 100 GB/s, then doesn't that imply people can generate 100 GHz signals to feed laser diodes, which is a high enough frequency to use waveguides?

Yes, that's true, but that's part of why fiber is more expensive for short connections; if you're going to use that kind of electronics you sort of might as well use fiber. Also, those use interleaved DACs with low efficiency, which is compensated for by the low losses in optical fiber. It's not uncommon for data cables to have 50dB attenuation, meaning that only 0.001% of the transmitted signal is received, and WiFi often has 60+dB attenuation.

But we can match the attenuation of typical fiber links with a 50+m waveguide.



Now, I said "5 GHz is about the limit for affordable hardware" and that's conventional wisdom that's guided designs, but if we're willing to accept lower output powers (<30 mW) we can use SiGe amplifiers at >50 GHz. [1][2]

Oscillators can manage higher frequencies than that. And we can use oscillators in place of amplifiers: if you inject a signal into an oscillator during startup, you can capture the signal phase. Then, you can quench the oscillator and do that again, receiving one symbol per quench cycle.

For transmitting, you can combine one oscillator per output phase, and quench undesired phases for each symbol.

The overall design might then be something like this:

data link

SRO stands for "super-regenerative oscillator". To get a suitable design for that, you could take this 180 GHz SiGe SRO, remove the variable frequency tuning, and scale it down from 130nm to a modern 55nm SiGe process.

Using this design, net energy usage of 6 pJ/bit should be feasible with current technology, which is less than PCIe typically uses.

The above picture shows a 20 GB/s configuration, but a 5 GHz quench rate for 40 GB/s should be feasible, and 10 GHz might also be possible.



RF switches typically handle lower frequencies than 175 GHz and have much lower switching rates than 2.5 GHz. The problem is that at 175 GHz, fast transistors have enough capacitance that they conduct the signal even when off. It's possible to do better than a general-purpose switch here: the switch frequency is fixed, so a resonant system can be used, and good isolation is not needed. An RF power divider would work instead of a switch, but it should be possible to get somewhat lower attenuation and noise than with a passive power divider. Here is a suitable switch design using 45 nm CMOS SOI.



How big would waveguides for this need to be, and how long could they be? That depends on the mode used.

To be practical, cables need to be flexible. So, the TE01 mode (with a circular waveguide) is one option: it has low losses and makes constructing flexible waveguides much easier because no axial conductivity is needed. A TE01 waveguide can be a stack of conductive rings, which can bend relatively well. But it has a higher minimum waveguide diameter than some other modes.

What if we used corrugated waveguide? Does that work? Actually, circular corrugated waveguide with the right corrugation spacing and depth is better than smooth waveguide. The corrugations make the HE11 mode possible, which usually has even lower losses than TE01.



The HE11 mode can't be made directly, but that's not really a problem. For a 175 GHz signal, the system would be something like:
PCB antenna <-> TE11 mode in 2.1mm ID smooth circular waveguide <-> mode converter <-> HE11 mode in 3mm ID (including corrugations) corrugated waveguide

Copper, aluminum, and silver-plated steel are all fine for this type of waveguide. Mode converters between TE11 and HE11 are relatively simple and have ~99% efficiency.

A 3mm corrugated waveguide should be good enough for 100m links, and larger ones could support longer distances; a 30mm waveguide should support 6km links.

I think that a single type of magnetically attached hermaphroditic connector is ideal, and the success of USB indicates to me that connectors should support DC power delivery. 48V is a standard choice because it's relatively safe and can be handled by low-cost MOSFETs. Cables wouldn't all need to have wires for power, but I think it should be an option on connectors.

So, a connector could look something like this:



Magnets are good because relying on friction wouldn't give good waveguide contact, but magnets of that size wouldn't prevent waveguide separation by the force of a long dangling cable. It should be possible to resolve this by having the inner part (with magnets and waveguides) floating inside the outer rim. The thicker areas of the rim are prongs, and friction of those would prevent pullout while the magnets would maintain waveguide contact within the slightly bent connectors. The details of this are left as an exercise for the reader.



Does this type of system have any advantages over optical fiber? Potentially, yes.

Fiber couplers are relatively lossy and expensive, but waveguide cable sections could be easily connected together with minimal losses.

Because this would require a relatively small amount of SiGe area, it should be cheaper than fast fiber transceivers. This might need 1mm^2 of 55nm SiGe per waveguide, if the transmission line components were on a 130nm Si interposer. The manufacturing costs for such a transceiver might be something like $3 for a bidirectional 40 GB/s link.



While dielectric filled waveguides generally have much higher losses, using them for a short distance to connect to the transceiver is better than using a longer wire to the coupler. By filling the waveguide with glass, sintered alumina, loose alumina powder, or alumina filled teflon, it should be possible to reduce the inner diameter to 1mm. That's small enough to place the waveguide ends directly against the top of the Si interposer.

back to index