> how do they get a reasonable noise figure? OSI is 8k for the 5534, but source impedance is around 400 ohm for 24 channels - way to low even after step-up...
5534 doesn't suck at 400Ω. Or really 1K6. Self-noise of 5534 is IIRC around 0.7uV. Self-noise of 1K6 is around 0.6uV. And while sum-amps are noise-sensitive, they should not be noise-critical. When you have good amplifiers, your noise figure should be set at the passive transducers (mike capsule), and everything else worked at high enough level that it won't degrade the noise spec. Figure mike-amps run at gain of 40dB, a 16-in mixer has 24dB loss, the level at the summing amp is still 16dB above the level at the mike, so even a somewhat loose noise level won't degrade the mike noise.
In this case, while the bus-amp sets the noise with all faders full-down, at any practical setting of the faders the channel noise will meet or exceed the sum-amp noise. That's part of why this card has all those gain-set options: to optimize the system gain-structure for best dynamic range.
> OSI is 8k for the 5534
But when you can't make the source fit the "perfect" OSI of the amp, you need another number. What is the -range- of resistances where noise is "low"? In BJT amps, this is really the Beta of the input device. Say the 5534 inputs have Beta of 100. The OSI of 8K really is just the midway point between an equivalent input noise-voltage (series) resistance like 800Ω and an input noise-current (shunt) resistance like 80K. If the source is 800Ω or 80K, the noise figure is 3dB. Splitting the difference gives even lower N.F., under 1dB. But in many situations, a 3dB noise figure is not a major contributor to total system noise.
> explain me the nasty feedback loop of this output-circuit? It looks like the generate negative impedance to cancel the primary resistance.
That's one way to look at it.
Consider the basic virtual-ground summing amp. The feedback works to keep the amp input at zero volts. When the summing resistors inject current into the summing node, the feedback has to supply a cancelling current to keep the input at zero volts. The resulting output voltage (forced to supply the canceling current in the feedback resistor) is the sum of the inputs.
In this case, the feedback extends into the transformer. It works to keep the flux in the core at zero, or as low as possible. This makes the input winding voltage near-zero. This also makes transformer distortion near-zero, while giving good summing action and transformer isolation.
> "Kurzschlussausgangsscheinwiderstand" stays for the impedance (@ 50 kHz in this case) one "sees" at the output if we short the primaries.
Leakage Inductance plus resistance and some other small details.
> is it possible to go with unbalanced summing or just bal. works?
As in any floating input: ground one side of the transformer and drive the other side.
But that loses the main benefit of this scheme. What is "ground"? Ground is a 6 foot long wire with many branches and many different signals leaking into it. Even if ground is a fat copper bar, it won't have the same voltage at every place. Channel 1 outputs a signal relative to the left end of the ground bar, but the summer references the right end of the bar, after channels 2-16 have dumped their own signals into it. That causes cross-talk in a large console. Also some ground returns are not even dumping clean signal: class AB outputs make big garbage in ground.
With balanced summing, the output of each channel is not referred to some "common point" that is really a dumping-hole, but gives us two terminals with the signal expressed as the difference of those two voltages. The summer takes the sum of the differences of all channels. "Ground" does not matter at all, or only as a minor leakage term (like -50dB).
Unbalanced makes perfect sense in smaller systems. Each channel is much simpler. None of these 4-gang panpots, or dual differential booster-amps after an unbalanced panpot. With careful layout you can run 8- and 16-channel 1-man mixers perfectly well, especially if you do not have a German Engineering Department's obsession with specifications. In rock-n-roll, a little cross-talk (like -50dB) goes completely unnoticed.