The cards are full length ISA, 2 Mbps, full power (100mW - the max allowed under many national communication commisions), and (in the version we used in Canberra) had only one channel. They did offer AntennaDiversity, though only with their own antenna. There was but one plug, which we had identified for us as some weird custom modification of a 75 ohm SMA-type shroud with a larger (smaller?) center pin making it a matching 50 ohm connector. Some people blew it away, replaced it on the board with a type N, and used it that way. Most people made pigtails from the original 2m-long white-tile antenna. The white tile antenna was two crossed dipoles on a printed circuit board. There were some embedded resistors and a diode on the board. What the card would do was bias the outgoing microwave signal either positive or negative, and this would select which antenna was used. Normally it would autoselect based on signal strength, but you could (usualy for white-tile target practice) set it one way or another. Our custom high gain antennas had no diode or resistors, so they had no such diversity. The one attempt to build such failed (well, it radiated, but not as uniformly as a dipole..) The receiver was very deaf by modern standards (-75 dBm if I recall what we were told correctly. That roughly matched our guesses at our own cabling losses, antenna gains, and signal strengths).
With 100mW and a homemade reflector you could slowly (over weeks) torch the white tile. If you left it up a pole during this time you would initially attribute the damage to lightning, until you repeated the experiment at ground level inside a metal shed. It's probably just as well the shed was cold and I spent little time in there while the second experiment was running.
The ISA card actually had a small waveguide built into it, the resonator and waveguide were what made it full length. From the software driver design (not by us) it was apparent that the card was just an ethernet chip attached to a DSS modulator and then attached to a microwave radio. You could turn the radio on and off, and there were two extra bytes in each packet for the network ID.
There was no built-in retransmission system or ACK's or collision avoidance - if we ever ended up with congestion (a problem we were looking forward to encountering 
) we had some pretty cool ideas for token passing and shorter-than-allowed-on-an-ethernet packets to share bandwidth in a hidden node environment. For some practical arrangements, the hidden nodes actually created more available bandwidth if you could time things correctly (which you could in principle do because the cards were so basic, the signal went when the packet was sent by the driver). You could adjust the length of the preamble as well (from 2 to the standard 8 bytes). In a simple network with four linearly connected nodes 1-2-3-4, where only adjacent nodes are connected, nodes 1 and 4 could of course transmit simultaneously to nodes 2 and 3 respectively. What could *also* happen was nodes 2 and 3 could reply simultaneously to nodes 1 and 4, even though 2 and 3 had a connected bidirectional communications channel - as long as they aren't trying to communicate with each other they could transmit simultaneously.
