So far I have not actually measured how long a hysteresis-based relay latch could hold a state. So I'm doing it now.

I have set up a little circuit with a RES15 relay (36 Ohms), a matching series resistors (39 ohms), a capacitor to set the state, and a LED to show the state.

The circuit is powered by a digitally controlled PSU set at 2.8V (mid-way between the 2.1V release voltage and the 3.62V latching voltage). After a while, the circuit draws 33mA (some heating occurs in the coil).

I have no idea how long the circuit can stay latched so it's not possible to use my multimeter (it would go into power-saving mode after some minutes instead of beeping). I have no timer either, so I connected an LED that would light up when the relay is released... and I count time manually :-D

Unless the register set is held in standby state during a debug session, I can imagine that a register would be toggled at the very least once a minute and the experiment is running for 3 hours now. The possible cause of perturbation in this test would be the poor wiring quality on the solderless breadboard, so I stay away from it to prevent any minute wiggle.

---

Retention of 1 bit requires 92mW so 64 bits (the YGREC8's register set) would draw 5.9W alone... This register set needs a separate power supply that is very stable and well filtered : 2.8V (5%) at 2.1A. Early experiments (with the YGREC16) have shown that the system is not stable if the latches' supply is shared with other circuits, which create a lot of switching noise.

I don't want to use a modern PSU so I'll go for the old good way : an AC transformer, a diode bridge and a large filtering capacitor. A very big capacitor is not difficult to find (10000µF at least, more is better to keep the ripple as low as possible), the diode bridge is possible (there were selenium rectifiers in the 1930s, each plate stands about 20V)  but a low voltage transformer is a different story today. I'm not sure I can find a transformer that can provide 3VAC under at least 2A (more is better). At least, I can finely adjust the input AC with an auto-transformer.

"Back in the days" when tubes/valves were kings, radio sets would provide a low voltage, "high current" output to power the heater(s). This is something to explore but what I have seen so far is 6V or 12V output, not 3V and only 2 wires.

Another alley to explore is partition : there are 8 bits that can be written at any moment (plus PC) so 8 subcircuits (one per bitslice) are possible, each with their own power source. Partition would be along the bitslice, not per register, because each write would create noise on 8 bits simultaneously and the strategy is to spread those spikes evenly. So each bitslice would have at least a local filter (capacitor + inductor) that can provide a clean power at 270mA. A local diode can also drop the current if needed.

If each bitslice has a local power input, the transformer can be partitioned into multiple smaller transformers. The bridge rectifiers become smaller too.

If I can find suitable 3V AC transformers, then adding the drop of a silicon diode bridge gives the right voltage : 3×1.4=4.2V, 4.2-(2×0.7V)=2.8V

I remember some old 1A bricks with a selectable output voltage : 3V, 4.5V, 6V, 9V, 12V. Inside, a PCB holds four diodes and a 1000µF capacitor, and the secondary windings were probably multiple 1.5V or 3V in series. I could rewire that in parallel to provide a stronger 3V output ...

---

test interrupted after 3h46m due to human error...

---

Test is restarted : 6 hours and no sign of weakness. Shall I call this experiment a success ?

---

In this circuit, the coil draws 33mA under 1.38V only. Half of the power is dissipated by the 39 ohms resistor (46mW). A different type of relay (with higher coil resistance and higher voltage) would decrease the overall current and ease the selection of the transformer.

The resistor is another problem and I wonder how to void it.

---

I stopped the test after more than 33h of flawless latching : the hysteresis is validated. A very stable power supply is the key!

For the power transformer, I just discovered the TSL series from INDEL. Two references (TSL40/001 and TSL100/001) provide dual 3.15V outputs at 3A each, that's about 18W (I ignore the very high voltage output). TSL100/001 also provides 5V and 6.3V at 2A each. The cheaper TS20/24 provides 9V@2A but 3V@1.5A only... So I ordered a TSL40/001.

What should be the maximum tolerable ripple on the power supply ? This relay latches at 3.6V and releases at 2.1V, the resistor of equal value as the coil nearly doubles the tolerance to ripple, and the margin is 1.5V. A ripple of 100mV or 150mV max. would be tolerable... But from experience : each relay has a bit of variation and this significantly reduces the margin of an overall group. This is explained there : https://hackaday.io/project/14628-ambap-a-modest-bitslice-architecture-proposal/log/45438-pre-biased-or-hysteresis-relay-logic

The 8 bitslices will have their own filter, to power 8 bits, and they can have a little resistor to adjust the current-voltage on the bitslice, so a batch of 64 relays can be "binned" into 8 lots of relays with very near characteristics. This was the purpose of #ReTest-RPi but I have not progressed on that front... Time to revive it !