In Dercuano I listed a bunch of “jellybean” FETs in 02017, coming up with this table:
| PN | Vds | A | ohms | Qg (nC) | ¢ | W | type |
|--------------+-----+------+-------+---------+-----+-----+------------|
| 2N7000 | 60 | .2 | 1.9 | 2 | 36 | .4 | |
| 2N7002 | 60 | .115 | 7 | 2 | 38 | | |
| IRF630 | 200 | 9 | .4 | 45 | 86 | 75 | |
| IRF9630 | 200 | 6.5 | .7 | 29 | 151 | 74 | P-chan |
| IRLI630G | 200 | 6.2 | .400 | 40 | 229 | 35 | |
| IRLML6344 | 30 | 5 | .029 | 6.8 | 36 | 1.3 | |
| IRLML6402 | 20 | 3.7 | .065 | 12 | 40 | 1.3 | P-chan |
| EPC2036 | 100 | 1 | .065 | .910 | 97 | | GaN |
| SI3483CDV | 30 | 8 | .034 | 11.5 | 89 | 4.2 | P-chan |
| FQP27P06 | 60 | 27 | .070 | 43 | 134 | 120 | P-chan |
| NTD4906N | 30 | 54 | .0055 | 24 | | 2.6 | obsolete |
| IRF7307 | 20 | 4.3 | .140 | | 83 | | dual (P&N) |
| BSS138 | 50 | .200 | 3.5 | | 24 | | |
| CPC3703CTR | | | | | 70 | | depletion |
| 2N5457 | 25 | .01 | | | 230 | | JFET |
| 2N5458 | 25 | .01 | | | 230 | | JFET |
| SiS410DN | 20 | 35 | .0048 | 41 | 94 | 52 | |
| PSMN4R0-40YS | 40 | 100 | .0056 | 38 | 88 | 106 | holy shit |
| IRF540N | 100 | 33 | .044 | 71 | 145 | 130 | fuck |
| IRF9540N | 100 | 23 | .117 | 110 | 189 | 110 | P-chan |
| IRF9530 | 100 | 12 | .300 | 38 | 138 | 88 | P-chan SyC |
One heinous Python expression later and we have them ranked by watts per cent:
>>> csv.writer(sys.stdout).writerows(sorted(
((float(v[1]) * float(v[2]) / float(v[5]), v[0], v[1], v[2], v[5])
for v in [line.split('|')[1:] for line in t.strip().split('\n')]
if v[1].strip() and v[5].strip()), reverse=True))
45.45454545454545, PSMN4R0-40YS , 40 , 100 , 88
22.75862068965517, IRF540N , 100 , 33 , 145
20.930232558139537, IRF630 , 200 , 9 , 86
12.16931216931217, IRF9540N , 100 , 23 , 189
12.08955223880597, FQP27P06 , 60 , 27 , 134
8.695652173913043, IRF9530 , 100 , 12 , 138
8.609271523178808, IRF9630 , 200 , 6.5 , 151
7.446808510638298, SiS410DN , 20 , 35 , 94
5.414847161572053, IRLI630G , 200 , 6.2 , 229
4.166666666666667, IRLML6344 , 30 , 5 , 36
2.696629213483146, SI3483CDV , 30 , 8 , 89
1.85, IRLML6402 , 20 , 3.7 , 40
1.036144578313253, IRF7307 , 20 , 4.3 , 83
1.0309278350515463, EPC2036 , 100 , 1 , 97
0.4166666666666667, BSS138 , 50 , .200 , 24
0.3333333333333333, 2N7000 , 60 , .2 , 36
0.18157894736842106, 2N7002 , 60 , .115 , 38
0.0010869565217391304, 2N5458 , 25 , .01 , 230
0.0010869565217391304, 2N5457 , 25 , .01 , 230
That is, in theory, the PSMN4R0-40YS (unavailable in Argentina) is capable of switching 4000 watts on and off for just under 90¢, so it can control 45 watts per cent, while the IRF540N and IRF630 (available, even listed on MercadoLibre for 80¢ and 65¢) are almost half as good, switching respectively up to 3300 watts or 145¢ (02017 price!) and 1800 watts for 86¢. I probably should have also listed the popular IRFZ44N (55V, 49A, thus 2700W, 76¢ locally) or IRLZ44N (55V, 47A, 87¢, thus 2600W, logic-level threshold).
Six of these monster transistors, plus the appropriate drive circuitry to control them, give you a three-way H-bridge to control a multi-horsepower “brushless” motor. One may be sufficient for a multi-kilowatt switchmode power supply, though maybe running off Argentine 240VAC you’d want two or three in series.
And, for electrolysis they can potentially drive material removal or deposition with jitter under 10 ns and pulse times of 100 ns or so. (This is inferred from the IRF630 datasheet, which has apparently renamed “HexFET” to “STripFET”: “typ.” 118.5 ns reverse recovery time, 5.6 ns turn-on delay time, 2.6 ns rise time, measured with 4.7 ohm gate resistance and 10 V; oddly they don’t state td so it must be something terrible.) For pulses, all of these MOSFETs support even higher powers; the IRF630 is rated for only 9 A continuous, but 36 A pulsed.
Of course you’d have to run the electrolysis through a step-down transformer or SMPS if you wanted to deliver that kind of power in a useful way; 3600 A at 2 V would be a lot more useful than 36 A at 200 V, which would mostly just heat up the water. Such a transformer with 10MHz bandwidth might be hard to find.
The gate charge is “typ.” 12 nC, so delivering it in 10 ns would require driving the MOSFET gate with 1200 mA, which is I guess why MOSFET gate driver ICs and pulse transformers are so popular. Getting a 10-ns-rise-time edge through the rest of your circuit is also doable, but nontrivial.
Suppose we could deliver 3600 A at 2 V for 100 ns. That’s 0.72 millijoules of energy, a perfectly manageable amount for ordinary circuits, and correspondingly 0.36 millicoulombs, or 2.2e15 electrons, or 1.12e15 divalent cations, about 1.87 nanomoles; for copper that works out to be 119 nanograms, and for iron 104 nanograms, assuming perfect Faraday efficiency. That’s about a 30-micron-diameter sphere of either of these metals: visible, but barely. (It would punch right through aluminum foil, though.)
(In practice such high current densities would be prevented by the formation of an insulating salt film on the surface. Also 0.72 millijoules in 100 nanograms is 7200 kJ/kg, which is still plenty to vaporize the metal.)
If an electrolytic cathode is flying over a flat metal substrate at 25 m/s, like in a laptop hard disk (but full of water), 100 ns is about 2.5 microns. The 10-ns jitter guessed at above amounts to 250 nm of imprecision. If you were using this to record information, you might encode 4 bits into the delay before each new pulse, with an average of 180 ns per pulse and rest, giving a data rate of 22 megabits per second.
If you wanted to archive a 10 GiB ZIM file of English Wikipedia on nickel foil this way, it might take an hour or so. You might want to reduce the current so you wouldn’t be gouging huge 20-micron-deep craters in the surface of the metal that would be hard to tell apart; 130 mA for 100 ns would suffice to give you a hemispherical 1-micron-radius pit. Spacing tracks 2.5 microns apart would give you 400 tracks per millimeter, and 4.5-micron-long pulse-and-rest cycles would give you 889 bits per millimeter, so 2200 bits or 278 bytes per square millimeter, so, all in all, you’d need 39 square meters of nickel.
Consider instead the average material removal rate (or deposition rate), supposing we can step down an average of 9 A at 200 V to 900 A at 2 V; that’s about 4.7 millimoles per second, about 300 mg/s of copper or 260 mg/s of iron, supposing divalent ions and 100% Faraday efficiency in each case. That’s about 1 kg per hour.
However, 10-ns precision at 300 mg/s means 3-nanogram precision in how much material you remove. If that’s spread over a square millimeter at 9 g/cc, that’s an etching or electrodeposition precision of 0.3 nanometers, roughly one atom. If we step down to the kind of precision we need for optical systems of about 40 nm, that works out to about a 90 micron by 90 micron area.
So if you were using such a transistor to control the low-precision hogging-out phase of cutting a first-surface mirror, your kg/hour hogging-out process would hit its limit at 40-nm Z precision per 90-micron-square area. Of course, that assumes you’re using laser interferometry or something for positional feedback of the electrode.
Then, by turning down the current for a finishing pass, you could overcome that resolution limitation and get the mirror surface more precise, still at the same bandwidth of a few tens of megabits per second.
One of the more interesting devices that can be usefully controlled at 10 MHz or more is a piezoelectric actuator. These don’t require a lot of current but they do need relatively high voltages.