Adventures in matching smd jfets

[Luke51411]

I purchased 100 mmbf5457 from mouser which is a 2n5457 in a sot23 package along with adapter boards from oshpark. I started out with 30. I used RG's jfet matcher to measure Vgs. I haven't seen an threads posting results though I heard that the smd fets are closer tolerance so I figured I would post my results. From a group of 30 I got 5 pairs matched within .003. There were no groups of 4. All of them were between -.846 and -1.222 with most being between -.969 and -1.197. They seem to be more consistent than the TH fets I've matched previously though I was hoping for a set of 4 in that group of 30. Here are the results:

Sorry for the image, I can't currently connect to the internet on my computer to copy and paste. I just ordered 70 more adapter boards and will post results upon matching the rest if there is interest.

[duck_arse]

 so's I don't sit staring at those numbers for an hour, what was the smallest spread for a set o' 4?

if you've identified each seperate fet, you can sort on the voltage and then stick the speadsheets #'s into the graphing prog that is probably included with, and then see how far apart the matches are.

[Luke51411]

I don't have any sets of 4 within .010. A couple are at .011 for the two farthest apart in the set. I imagine I'll have quite a few more matches wen I do all 100. I was just curious if they would match up easier than the TH version or not.

[Davelectro]

They look pretty consistent to me. That's great.

[Luke51411]

Yeah, much more consistent but I was still a bit surprised there weren't more close matches with how consistent they are as a whole. It will be interesting to have a larger sample. Just out of curiosity, would there be a way to match them without installing them on an adapter board? For use in a fully smd build for example. I know you can get a matched set in a single package but it seems like it would be difficult to match them by hand for larger matched sets. I haven't delved into the world of smd yet at least beyond adapter boards but I imagine I will eventually.

[Jdansti]

What comes to my mind is how reproducible are the measurements (precision)?  So if you remeasured the four that had the closest readings, what would you get if you remeasured those same four a second or third time?  If your measurement system is giving you a spread, it might be best to make multiple measurements, average each JFET's readings, and use the averages.

[Luke51411]

That is a good point John, I'll try it out.

[R.G.]

I was about to post a note on the issues of accuracy, precision, and repeatability, but John beat me to it.

0.003/1.1 = .00273, or .273%.  What is the accuracy spec on your voltmeter?

Frankly, using a home-made lash-up for measuring, getting down to 1% accuracy is something of an achievement. I would be happy with quads varying no more than about 0.010.

Here's a good quick read on the differences between accuracy, precision, and repeatability. It's quick, and serves as an intro to some of the issues.
http://www.ndt-ed.org/GeneralResources/ErrorAnalysis/UncertaintyTerms.htm

The other bit of uncertainty you get into is that the quick and dirty matcher only selects one point on the operations of each JFET.  It is entirely possible that they may match at one point, but not at another in operation. Since the idea behind a quick and dirty FET matcher is to get FETs from a commercial batch that will match OK and produce decent phasing, there's a great exposure to diminishing returns in measuring them ever more accurately. The equipment to do this reliably is expensive.

[Luke51411]

So basically .010 may be more accurate than the matcher can reliably measure. I thought about building the improved jfet matcher but I wasn't quite sure what that was measuring and what to look for. For me I guess id like to just have a fair amount of roughly matched sets and try them out to see how they sound.

[R.G.]

It's not necessarily the matching fixture that's the problem. The matching fixture, however you build it, has an internal set of stuff that results in a "reference". This sends a measured amount of current, voltage, etc. into the device under test. You then measure this result with your meter. The meter itself has its own accuracy, precision and repeatablility. Many meters specify in the small print "accuracy: 2% (or 1%, or something) of reading +/- 1count." Digital meters (and I think you used a digital meter, I've never seen an analog meter I could read to four or five significant figures" have their own issues internally.

So you have a "reference" in the lashup that has no particular accuracy, except that for over short periods of time and without changes in temperature, humidity, etc, it treats each device the same. It has poor precision (you don't know exactly what conditions it's providing right now), essentially no accuracy, but excellent short term repeatabiliity. That is, it treats the next JFET almost exactly like it did the last few. Your meter is then what is doing the accuracy, precision and repeatability. Unless meters are frequently calibrated themselves, they drift, get scale errors, and so on. Your meter has better repeatability than the measurement lash-up, and probably better long-term accuracy, but it has its inherent errors on every single measurement. The long-term repeatability of the test fixture is going to be bad, as there was no care in the design done for keeping it stable.

So it's good for one thing - how far apart are these JFETs right now, and don't look too closely at the exact numbers. I don't remember the improved JFET matcher well, but it may or may not be better for repeatability and stability.

However, there is a better way to look at your numbers that is very instructive. Label your JFETs with a name or number each so you can identify individual devices. I suspect you've already done this. In a spreadsheet, enter each identifier and its reading. Then compute the average of the whole set, and the % deviation from that average for each device. If there are one or two that are dramatically different from the others, take them out of the calculations and recalculate the average and % deviation.

Now select for groups with similar devations from the average. This process "magnifies" the differences.

And don't get all wrapped up in thinking that hyper-precisely matched JFETs sound better. They might - or might not. And JFETs are a poor place to start looking for matching and precision. 

[Luke51411]

That makes a lot of sense. Thanks. So I could test them again at the same time if I get similar readings in the next batch to assure a decent match.

[R.G.]

You could. Not a bad idea. But don't be alarmed if the actual numbers are different. Something may have drifted with temperature or time on the test setup.

I would expect the relative differences from the batch average to be much closer than the specific numbers. At least for measuring sessions short enough that the supply voltage and part temperature in the test setup don't change much.

[PRR]

Almost any four consecutive parts on that list gives you 3% matching.

Which is probably better than your capacitors, and even many resistors.

[bool]

Next thing in the quest for the ultimate match would be the curves.

[electrip]

DCA 75 PRO

[Luke51411]

I'll try out several sets in a couple p90 style circuits tonight after work and post the results as always 

[R.G.]

Almost any four consecutive parts on that list gives you 3% matching.

Which is probably better than your capacitors, and even many resistors.

... and correspondingly makes me think that he accidentally got a whole batch off the same wafer!

[Luke51411]

Hopefully the rest of them will be grouped this closely as well!

[Luke51411]

I just tried a set of four in a p90 style phaser and it sounds great! There is more phase than my best matched set of 2n5457s and they were matched pretty closely if I remember correctly.