Works even better in 74CBT transmission gates...

But I've got a three rails of ten MY4N 12VDC QPDT relays to attempt a logic demo.
Sadly, the internal LEDs only light one direction and not terribly bright at that...

How a similar but not exactly identical ALU might appear as 6nS ICs...
Wired for Manchester Carry this time, omitted zero, and combined some controls.

If drawings failed to make the point, No Ripple.

A simple maze of switches can be setup by A vs B without waiting to know carry, borrow, or equal.
That maze quickly solves itself by path of least resistance.

250pS propagation delay per switch (74CBT3253). Perhaps a little more considering fanout to final XOR.
But 6ns to prefix those 5ohm switches occurs in one step of parallel time. Without a parallel prefix tree.

Where it breaks from Manchester: Generate isn't hardwired to VDD. Annihilate isn't hardwired to GND.
Instead, control lines LT GT allow new magnitude comparator functions.

Where it breaks from Konrad Zuse: Open circuit is not a valid logic level. Logic lows are driven.

Where it breaks from i8086, you would have to read Ken Sherriff's blog for that (different Ken).
http://www.righto.com/2020/08/reverse-e ... 8086s.html

You might be thinking "The above doesn't look much like i8086".
Not yet, but let's observe where this is going...

Mucking with this puzzle in attempt to include rotation I also let
XOR vs NXOR determine the propagation of CARRY vs BORROW.
This allowed to omit the ADD vs SUB control line.

Was about then I notice the other Ken had blogged a teardown
and partial reverse engineer of the i8086. Scroll down about 1/3
of the article to find his drawing "One bit of the ALU."

At least my way of rotating still looks different. RP is short for
"Replacement", the new flag to shove when propagation of a
prior carry/borrow/equal flag does not occur. Also a convenient
signal to intercept for rotations and reversal by RP loopback.

This would be my pile of junk to build the relay version.

Coils solve XOR in either direction. MOVs clamp flyback
+/- 18VDC in either direction. But they don't allow the
one-way LED built into the relay to work either direction.

Thus you see a bag of diode bridges to replace MOVs.

Simulation given with OP of this thread chose lamps
that don't care a direction. No bridges are needed to
solve XOR. Bridges also solve XOR, but redundant to
coil or lamp logic and not really the purpose.

Lamps... Maybe I don't need diodes or MOVs. As drawn in the sim, a low resistive lamp is bidirectional and should also damp coil flyback. Just learned Christmas surplus may be strung with 12V lamps and sockets, or lamp leads might screw down without sockets. Far cheaper than pinball bulbs. Wouldn't take much to outshine that dim one-way LED inside each relay.

How do you do a lamp test, to check for dead bulbs?

Ken, Is this your first attempt to build a hardware relay-based logic unit, or possibly you have other photos or designs that you put into hardware? Looks like really interesting stuff. Michael

Probably isn't any input condition that would light all bulbs at once.
Input conditions to test the Zero lamp have already been shown.
Two other settings should cover the rest...

Well, there was one-bit hot glued to a popsicle stick. Built from the trash bin.
Regularly had to thump the benchtop to unstick it. Using majority logic then,
rather than truth table logic now. So the relay count wasn't quite as efficient.
Wasted one extra DPDT, yet didn't offer quite as many functions...

Some of the sticking could have been micro-welds from undamped coil flyback.
By the time I found MOVs to deal with that, I had new relays and a better plan.

I look back at this old drawing with few clues what could I have been thinking?
Majority3 of X,Y,Z is Carry. MAJ3 of X,Y,1 makes OR. MAJ3 of X,Y,0 gives AND.
Something like that. Further specifics I've totally forgot. Not worth it.

"Easy to make MAJ3 from MUX4. Hard to make MUX4 from MAJ3.", or
"Easy to glue junk to sticks. Much harder to unstick."
Might be other morals to this story.

I would not suggest to build this variant. Shown only to answer a question.

Tiny bag of 100 "12V 3mm Grain of Wheat" bulbs have arrived. Smaller than Christmas tree lights. Leads shorter than expected too, but sufficient to reach the relay screw terminals. Strangely, leads penetrating glass proved tin-able (though only tried the far ends) with ROHS solder and water soluble flux. No difficulty attaching wires. Might try vs perfboard next.

5V@35mA glows well enough to be noticed, even in a bright room. But 12V@55mA is painfully bright, not to be looked at, perhaps to illuminate another nearby surface? Would need a darkroom to spot this bulb at anything less than 3V. 10V strikes a good balance. Just enough to activate the coil and swamp the dim useless one-way LED built into the relay.

After burn-in for an hour at the full 12V, the relay coil feels a lot warmer than the bulb. Touching glass, no hotter than an LED. Not likely to unsolder itself or melt plastic. Unfortunately not quite small enough to take the place of the internal LED. Will have to be external on the screw terminals.

Ability to subtract in either direction with either polarity of Borrow means I could pre-XORvert A instead of B. Now B gets 6nS time (in parallel with A's XORvert) to do something else. Maybe pre-ANDpass, ORpass, NANDvert, or NORvert? Update our spellchecks after this post, but the option to override and force B=00000000 or 11111111. Either might allow increment, decrement of A and other tricks without new delay.

But how/where an extra 74LVC1Gxx gate might be shoehorned or deadbugged? My DIP28 adaptors offer slightly larger SOP pads to work with on the flip side, where I hide the local gate to gate wiring.

Before the above thought, I had re-imagined my solid state circuit as a borrower. To clarify sameness to the relay ALU.
Forgot to note +6nS delay twice for XOR gates. Though XOR could be only +1.3nS in 2.5V 74AUC logic. I prefer 5V.
Those extra delays aren't in the critical path and don't ripple.

I prefer T11 tied to ground and combine T01 with T10, reducing spaghetti by two control lines omitted.
Can't remember now why that trick made me want to wire for carry. Should also work with borrow?

The attached drawing is borrow. And for clarity, no control lines omitted.
Already see several places this drawing could be made better. No time for it right now...

Suppose B were replaceable by the totally daft constant of 2?
Together with Carry/Borrow's ability to add or subtract 1
Can I now increment and decrement by one, two, or three?

Or instead drive B to zero for conditional addition?
Might help multiplication, but putting the option to skip
elsewhere and avoid the do-nothing add might be faster.

Or better just specify the entire byte B as it was before?
Yeah, I'm thinking these tricks might just be nonsense.
Not every idea goes somewhere, but sometimes need
drawing before the flaws become apparent.

Not get carried away with unnecessary feature creep.
The minimum requirement is to control program flow
and concatenate arguments to lookup table addresses.
Shadow tables to faster SRAM, and bootstrappy stuff
like that...

Everything non-essential to lookup should be tabled.
Though this ALU is twice as fast as 35nS MRAM,
it's probably not faster than 10 or 12nS SRAM.
Never compute anything you can lookup faster.

Omitting all diodes and let incandescent bulb damp the flyback.
-18V spike still a hazard for ICs, but harmless to other relays.

I wrote -12V in error earlier, only noticing/fixing on 20230828.
Scopeshot ground was offset one glidline above center.
Still better than my 18V MOVs wouldn't clamp till +/-24V.