That is reassuring but I cannot see being able to duplicate that performance without hybrid or maybe waveguide construction.
I know I'm being repetitive ... but one of the edges on the scope trace of the pulse using the reverse recovery method dropped 80-20 in 4 picoseconds and he just used CPWG on a normal PCB with an end launch connector. It would require a fair amount of experimentation to get it to work though, you can forget about trying anything out in spice.
I mean duplicating the performance of an SD-24 which couples the edge into the 50 ohm termination (25 ohm impedance) of a sampling head without disrupting the sampling process and without using a power divider. The pulse generator is effectively transparent before sampling occurs.
The Tektronix 067-0587-02 and 067-0587-10 1 GHz standardizers generate differential edge rates better than 150 picoseconds good enough for 1 GHz oscilloscope calibration and may give some idea into how they went about it in the SD-24.
If you just want something simple why not just drive a RF NPN into saturation without emitter degeneration with relatively fast logic? (ALVC for instance.) This should give you a damn good edge enhancement all on it's own ... it's not like these RF BJTs are slow even when used normally, cheap too.
The transconductance based reference flat top pulse generators I have studied work in the way you describe but without saturation and so far the fastest I have seen is 150 picoseconds using an integrated end-fire configuration like you see in the fastest integrated ECL/CML parts. A PG506 uses discrete through-hole 2 GHz PNPs and 4 GHz NPNs to generate 600 picosecond (10 to 90) negative and positive edges but I believe performance is limited by layout considerations. The similar National Bureau of Standard design is the same speed and the Picoseconds Labs improved implementation is 425 picoseconds.
I want to try this without saturation using a transconductance and diode switch based design. Where it gets tricky is that the straightforward general purpose implementations have some complementary elements and PNP RF transistors are slower. Lead inductance is a problem which Tektronix often solved by using dual base, collector, and emitter connections but nobody makes parts like that.
I have not exactly given up on step recovery diodes but I have decided to grade my own from varactor and PIN diodes. I will have to add transistors as well now based on that link to the NPN based step recovery design. Tektronix used all kind of odd diodes in step recovery applications.