Every switch matrix spec sheet I get asked to quote lists port count and insertion loss. Almost none lists a leakage budget. That is the number that decides whether an 8x8 matrix actually behaves at 40 GHz, and it is the number that is hardest to fix after the chassis is built.

Here is how the three common architectures differ, and where each one quietly fails.

Three ways to build the same matrix

Fan-in tree. Each output port gets one SP(N) switch, and its throws land on the N input lines. One switch in the path, so loss is the lowest of the three and isolation is a single switch's off-isolation. It is non-blocking only for one-to-one connections: if input 1 is feeding output 3, nothing else can use input 1 until you move that switch. Cheapest, lowest loss, least flexible.

Full crossbar. A splitter on every input, a switch on every output. Any input can reach any number of outputs at once, which is what you want for a parallel ATE rack or a redundancy test. The cost is the splitter: a 1:8 resistive splitter burns 9 dB before the signal ever reaches a switch, and a Wilkinson version burns 9 dB too plus makes the unused branches reflective.

Two-stage cascade. Fewer switch throws, more routing options, at the price of two switches in series. Loss roughly doubles in dB, switching time doubles, and the isolation arithmetic gets interesting in a way that usually does not go the way the designer expects.

Isolation adds in dB until it stops adding

Two 60 dB switches in series look like 120 dB on paper. In a real chassis you will measure somewhere between 85 and 110 dB and then stop seeing improvement, because the leakage floor is no longer the switch. It is the path around the switch: connector shells bonded to the same panel, the gap between two SMA bulkheads 15 mm apart, a ribbon cable crossing under the cavity, the ground plane of the driver board.

The practical consequence: if you need more than about 100 dB, do not buy it by cascading switches. Buy it with packaging. Separate cavities, feedthroughs instead of bulkhead connectors, absorber on the lid, physical distance between the input and output connectors, and a driver board that is not sitting under the RF path.

Also check isolation at the frequency you actually use. A switch quoted at 80 dB at 1 GHz is often 50 to 55 dB at 40 GHz, because the off-state capacitance of the contact and the coupling across the cavity both scale with frequency.

The number ATE engineers should be asking for

Absolute insertion loss is calibrated out on the first run. What does not calibrate out is the spread of insertion loss across paths, and how much that spread moves after a few hundred thousand cycles.

Ask for three things instead of one:

  • Insertion loss spread across all paths, worst pair, at the top of the band
  • Insertion loss repeatability on a single path over 100k operations
  • Return loss spread, same conditions

A matrix with 2.0 dB typical loss and 0.15 dB spread is more useful than one with 1.4 dB typical and 0.6 dB spread, because the second one forces a per-path calibration file and re-verification every quarter.

Terminate the ports you are not using

An unterminated reflective SP8T output reflects the leakage back into the cavity. In a crossbar the effect shows up as ripple in a path that has nothing to do with the unused port. Internally terminated switches cost more and add a little loss, but they make the matrix's behaviour independent of which ports happen to be idle, which is exactly what you need when the test sequence changes every product revision.

Latching relays, not failsafe, for large matrices

This is the failure mode that kills matrices in the field. A failsafe relay drops to a default position when the coil loses power. Put 128 of them in a chassis, lose the supply while the source is live, and you have hot-switched 128 contacts simultaneously into whatever load was there. Contacts pit, insertion loss drifts, and the failure looks random for months afterwards.

Use latching relays, drive them with a sequenced controller that never changes state while RF is present, and put a hardware interlock on the source. Also size the supply for the inrush, not the average: 128 latching relays at 28 V pull real current for a few milliseconds each, and a supply that sags during a re-route will leave relays in an undefined state.

Before you send the RFQ

Have a number for each of these:

  • Port count, and whether you need one-to-many or only one-to-one
  • Highest frequency, and isolation required at that frequency
  • Insertion loss spread target, not just typical loss
  • Switching time including driver and sequence
  • Lifetime in full re-routes, not individual switch operations
  • Control interface, and what happens on power loss
  • Connector series and the number of mating cycles the bench will do

Where I sit

I work on these parts at DB Design / MeiXun in Wuxi, so read this as notes from the manufacturing side rather than a neutral survey. We build electromechanical coaxial switches from SPDT to SP12T and DPDT rated to 2 million operations and 67 GHz, waveguide switches to 110 GHz, and PIN diode switches for the cases where the re-route has to happen inside a pulse.

If a matrix in your rack has started drifting, the leakage budget and the relay type are usually where the answer is.