.
Reading today about the Xerox Alto, a microcoded machine which can emulate something a bit like a Data General Nova, both being 16-bit word-oriented machines with multiple registers, I came across this examination of quite a few historical machines with 16 bit wide instructions.

I see, for example, that the TI-9900(*) has 4 bits to specify each of the source and destination registers, a bit like our OPC-5 machines, but where we decided to set some bits aside for a predicate (to make conditional jumps easy) they decided to use bits to select between addressing modes.



(*) The TI-9900 didn't do so well in its home computer usage, in the TI-99/4, in part because the 16 registers were in RAM and in part because most of the RAM was further away, the far side of the video chip. But if the 16 registers had been on-chip as originally envisaged and if the main RAM had been as large and fast as it was in competitor machines, it might have been a different story.

Very local though.. a Soviet machine mentioned in passing, otherwise only US computers as far as I can tell.

The NORD-1 from 1967 was also a 16-bit computer, the later NORD-10 and NORD-100 too, they were instruction compatible but introduced variations which didn't fully follow the initial NORD-1 pure format, as the years went by. I can't make drawings such as in the article, but a 16-bit memory reference instruction had an op code in the 5 leftmost bits (15-11), followed by 3 address mode bits (10-8), followed by an 8-bit displacement (7-0). Depending on the address mode, the address would be relative to the program counter or to the B- or X-register.

There were other types of instructions too, e.g. shift instructions and register instructions, and more. But they too had the op code in bits 15-11, the rest was interpreted differently.

An interesting machine. I saw the other day a figure, that adding one flip-flop to one of these early computers would raise the price by $300 IIRC. That gives us some hint of the pressure to make best use of every bit, and by extension every logic gate. I can't quite put my finger on the relative importance and benefit of, say, addressing modes versus ALU dexterity, but I'm sure there's something there.

And of course, that journey from accumulator-only machines, to those with index registers, and then a register file, is very much an economic and technology journey.

Found the reference: here.

It's hard to imagine that a flip-flop could cost that much when a vacuum tube was only a dollar or so.

It might be that a tube for computer purposes needs different characteristics and is produced in much lower volume than a tube for audio or radio purposes. I did read something about some component which is normally optimised for linearity - for all analogue applications - and which was just about usable for digital purposes. I think also a flip flop contained a handful of components. Here's a flip flop from EDSAC:
http://www.billp.org/ccs/Edsac/Logic/Lo ... odule=CCU9