Perhaps if you had an 8086 / 80286, you could only access register-register operations and the register-memory operations came later.

Yeah, found that one too. That post refers to a 1167 "chip", for which I never found a reference. Instead, the 1167 is supposed to be an "L-shaped PCB with the 1163/1164/1165. The compaq PCB seems to be a proprietary implementation to get the Weitek processor hooked to the frontside bus, the ISA part is just for power.

I had the Compaq Deskpro 386/20 at some time, including the technical manuals. I can't remember that the board had two 50-pin connectors to match the sockets on the Compaq Weitek card, but maybe I missed them.

Otherwise, the card might have come with an interposer board between the mainboard and the 386DX to tap off the local bus.

pcjs.org has scans of the excellent technical reference manuals uploaded to archive.org for the Deskpro 386/20, which includes detailed descriptions of that card in chapter 14. See archive.org/details/trg-deskpro-386-20-vol-2-1987-10/page/n291/… and on

The manual says that the ISA connector is not even used for power, just for mechanical stability. Furthermore, the manual confirms what I expected about the 1167: "...the three chips (collectively called the Weitek WTL 1167)..." (on page 294 according to the archive.org viewer).

That manual also confirms the theory that there is another board to connect to the 386 local bus. The Deskpro 386 obviously has a 3167-like socket which is a 387 socket with an extra pin row. As the 3167 was not a thing yet, the layout might be fully incompatible, or Weitek might have designed the 3167 specifically to fit the same socket as the interposer board. Mechanically, they look compatible. See page 309 for the socket type and the pages before that for pin assignments.