Improving an LM318 Op-Amp Follower by Prototyping Techniques

LM318 schematic for fast follower operation

This little exercise came from trying to make a differential NTSC video sender over CAT5 twisted pair from discretes (I know - don't do this at home, just get an IC that does this for you like the MAX4447 / MAX4146 pair).
Throughout this article I will be making liberal references to the online free book "Op-Amps for Everyone" Texas Instruments Document SLOD006B by Ron Mancini.
The Op-Amp I decided to base the design around (again there are discrete transistor versions of this circuit that are more capable than this) is the incredibly venerable LM318.
Its main specs are:

• Small Signal Bandwidth . . . 15 MHz Typ.
• Slew Rate . . . 50 V/µs Min
• Bias Current . . . 250 nA Max (LM118, LM218)
• Supply Voltage Range . . . ±5 V to ±20 V
• Internal Frequency Compensation

 [The LM318 is a pretty incredible design for its age - it really is! Especially since high performance PNP transistors were simply not available. For more details please see p19 of "IC Op-Amps through the ages" by Thomas H. Lee, 2000 rev. 2002]

The circuit I am trying to build is from the datasheet as a "fast follower" using lead compensation (SLOD006B section 7.6) and built on a standard prototyping board (called a breadboard at times).

Attempt 1: Standard Prototype Board

[Note: Bypassed both supplies with 0.1µF and 1µF ceramic capacitors.]
After building the circuit it was stimulated with a 1V Pk-Pk sinewave swept from 1MHz to 20MHz, and we should see the perfect flat gain of 1 till it rolls off at approximately 15MHz. Of course, we don't!
The gain peaks at approximately 2MHz at about 3 and then crashes to zero. The good news is that I have seen this behavior before and it is particularly well explained by SLOD006B section 7.7. Stray capacitance on the inputs. On a standard breadboard the stray input capacitance between V- and V+ is a minimum of 2pF. A good discussion on this can be found on EEVBlog #568.
First - to the simulator! Now SPICE is brilliant. Models of Op-Amps are not, and vary from modelling just the gain and frequency rolloff to full with major parasitics and noise sources depending on part and manufacturer. However if the SPICE model behaves the same with appropriate parasitics added it can confirm the behavior:

SPICE Frequency Response of the LM318 "Fast Follower" with a 10pF parasitic wired between V+ and V-

 This is a pretty damn good match to the observed behavior.

Attempt 2: Partial Air-Wire

[Refer to Linear Technologies App Note 47 p27 - a legendary app note, you should read the whole thing!]

Partially Air Wired LM318 "Fast Follower" with V+ soldered directly to the input 10k resistor and compensation pins lifted

The results of this approach was a massive improvement. The peak in the gain has smoothed out and I can sweep the stimulus (1V Pk-Pk sinewave) from 1MHz to 11MHz before the gain is 50%. The gain now peaks at approximately 2 around 6MHz.

Attempt 2: Full Air-Wire

In an effort to eke out further improvements the circuit was built fully air wired. I cannot emphasize the importance of removing ALL traces of glue and flux from around the pins of the Op-Amp.

Fully Air Wired LM318 "Fast Follower". [Note that the test configuration did not use the yellow crocodile clip lead to inject the signal and all grounds were brought together in the loop at the top right.]

Results from this approach were a further improvement of approximately 20% with the gain peaking at around 1.8 at 8MHz and the 50% point pushed out to 13MHz.

Thoughts

I have a feeling that to improve linearity it would be worth me building a circuit with a gain of 2V/V. From reading Op-Amps through the ages it seems that this design was not optimized as a follower and in fact the datasheet contains dire warnings of trying to build it without resistors and a compensation capacitor in the feedback loop.
I am continually impressed by the Op-Amp designers of old who were able to make these high performance circuits with pathetic silicon technology and little to no simulation support. I shall revisit this article and capture some more scientific measurements at a future date.