PretentiousFood

I don't think I've ever seen an SRM-700T in the flesh, but I've had pretty lousy experience with the Nexperia duals. I tried the PMP5501 and NSS40300 matched pairs as input buffers to drive a differential pair, and measured a 1/f corner as high as 5kHz on the NSS40300. Both were too noisy and had audible tape-like noise. Used in a differential pair with this much gain, I'm sure the noise is much worse. 1/f noise is generally a result of wafer contamination at the fab; maybe these other parts are made on a cleaner process, but the Toshiba parts are very quiet. If we want JFETs, the JFE2140 is pretty compelling. Much tighter matching than the LSK parts, and a bit cheaper, too. https://www.ti.com/product/JFE2140

Have you tried measuring the noise from V+ to V- when the pair is stacked? I ran into trouble stacking low voltage Mean-Wells in this manner; not sure if it's the slightly different oscillation frequencies from unit to unit, the stray capacitance, or something else, but it's significantly worse than when measuring a single unit. A good workaround if this proves to be troublesome is to place a common mode filter on the output of the SMPS, and to stack the outputs of the filter rather than the SMPS V+/V- pins. This eliminated noise that differential mode filtering and regulators couldn't take out in a couple of my builds. I've found that I need a fairly large value CM choke and cap, at least 50mH for the choke and upwards 1uF for the cap. These Kemet parts have worked well for me. That said, perhaps these don't need them at all.

If you don't mind downloading a 25mb pdf, the National FET Databook actually shows the dies. Some aren't perfectly symmetrical, so while they'll work in either configuration, I'm guessing they could perform slightly differently if D and S are flipped (e.g. process 88 shows more surface area for the S than D-- might impact gm?). http://bitsavers.trailing-edge.com/components/national/_dataBooks/1977_National_FET_Databook.pdf

I would probably skip the feedback and cathode Rs (which reduce gain and increase rp), and use a lower mu, lower rp tube like a 12AT7. If you do want to stick with the 12AX7 and nfb, they really do benefit from a plate load several times greater than rp, which shouldn't be too hard to pull off with an lnd150-based gyrator. The DN2540 only really shines at quiescent currents above 10mA or so. Something like a BSS126/DN2540 cascode or LND150 cascode may be better for the tail of the LTP, but note that the LND150 usually cuts off around 700uA if you try to cascode it.

Neat! I'm not sure the feedback works there; the two outputs are summed through R22 and R23 and terminate to a common-mode node. But the two stages are cap-coupled, so common-mode feedback won't do much. Perhaps add some resistance between the two cathodes? Or connect the feedback resistor to the plates rather than cathodes for some trendy schade feedback? What's C2 for?

Don't do it Brent!

Mouser's finest! I'm not sure about the hfe grade as I can't find the receipt. The package says FCE. FCE is not an hfe grade, but E is the highest and sounds like something I would buy. Shoot me a PM if you'd like.

I think I have about 100 of each (ksa992/c1845) that I'll never use if anyone needs. Can check tonight.

If you are lucky, it gets a little bit better if you literally reverse the polarity, input to the plate, and take the output off the grid.

@cetoole was visiting, so we made a slightly absurd amp using parts from the bin. I had originally milled the chassis with the intention of building a 6E5P triode -> 4P1L triode SE; then realized that would be a bit awkward, since the 6E5P is rated for higher dissipation than the 4P1L. I may have also forgotten that 4P1L filaments need power supplies when I bought my power transformer, too, so that was a non-starter. Not wanting to leave two unused holes in the panel, the obvious solution was to throw some 0C3s in. Glow tubes are pretty; it follows that any amp that uses them will sound better than one that does not. A normal person might be tempted to string them in series and use them as a reference for a regulator. Instead, each channel gets its own shunt regulator with a voltage reference made up of a zener stacked on top of the glow tube, and a SiC shunt device. R7 ensures that the tube strikes. And the current source (which is nothing inspired) The regulator terminates into the cathode rather than ground, keeping the cathode regulator out of the output current loop. It also forces the PNP to run at constant current regardless of cathode current, meaning its impedance will also be constant. It sounds like a single ended triode (i.e. terrible). But it was an awfully fun build. Slightly tempted to parafeed or otherwise upgrade the iron. It's surprisingly silent when nothing is playing.

No access to Spice right now but I've had great luck with this circuit and a variation using two PNPs and two DMOSs. If base current starves the zener, use a Darlington or replace the LND150 with a BSS126 or BSP135, I think the reason why the DMOS circuit is so good is that it does not have a bias string that runs parallel to the CCS. The way it's cascoded also holds the lower device's Vce or Vds constant rather than holding Vbc constant, as most PNP cascodes do. Using a BJT as the lower device almost always results in more transconductance, and thus higher impedance.

What kind of noise? I've had good luck filtering switchers with a CLC with -3db a decade below fosc (so ~10khz), to pick up where the regulator starts to drop off. Gives values in the low mH and uF range.

It's going into a Parasound D/AC 1600. It's still a work in progress-- I needed to replace the DAC power supplies because the stock ones were remarkably derpy. Pulling the I/V stage causes the regulator output voltage to go up. I put in some shunt regs and ripped out approximately a million bypass caps that were to far from the load to do anything. I don't like GICs/FDNRs. That's what the Parasound had. The big problem I see is that the op amps don't have sufficient gain at frequencies that we care about (352.8KHz and multiples) to actually work. The PCM1702 specs this fellow, which has just 30dB of gain that rolls off very quickly at 300KHz+. Not only is the impedance driving the shunt op amps going to be sort of high, it will also change with frequency. I like Sallen-Key filters because they can be built around an emitter/source follower. Those are much more dependable. By cascading it with an RC pole, you can build a third order filter with just a current sink and a BJT as active devices, and it seems to work well into the lower MHz without doing anything weird.

It's my own, but very similar to @cetoole's. It's a folded cascode, with differential filtering and a discrete Sallen-Key filter. Trimmers instead of an op amp servo, and a nickel sandwich on the output. Using some nice TO220 parts that let me bias it very hot.

For my Parasound. Getting it to bias was like resistor sudoku.