# GHz class sampling oscilloscopes on the cheap

I was recently perusing Picoscope's line up of sampling oscilloscopes, as they are all way outside my price range, I'd like to have a go at putting one together as a "hobby" project. ("hobby" herein meaning non-commercial, not breadboards and duct tape)

I would like to know from someone more well versed in this area if my idea is even feasible before I go spending $500+ on components. The heart of the system revolves around a TI THS788 (a glorified 4ch stopwatch with an LSB of 13pS and a dedicated trigger input). Basically, the front end consists of several really fast comparators (most likely the HMC674LP3E). You set your trigger level on one comparator, which when triggered, starts the THS788 counting (or really sets 'zero time' as it's a free running timer). The other 4 channels each have their own comparators and associated reference DACs, you set each comparator to trigger on a slightly different voltage and the THS788 records the time delay from the trigger. (with all differential signal lines properly length matched of course, we're dealing with picosecond delays after all) By repeating this process over and over with slightly different setpoint voltages on the input comparators, you can build up a picture of the waveform (or at least it's rising edge). As the THS788 has 13pS time granularity, this would in theory give you an Equivalent time sampling rate on the order of 75Gsps (and the input bandwidth of ~10GHz on the comparators helps too). That'd allow for some pretty decent TDR measurements and/or eye diagrams provided that a fast enough pulse generator was available. So should I dust off my box of K connectors and 40Gig semi-rigid or am I getting ahead of myself? simulate this circuit – Schematic created using CircuitLab • Look at the design of many, many such devices before attempting to make your own. – Ignacio Vazquez-Abrams Nov 8 '16 at 23:26 • Are you trying to make something like a Wavecrest analyzer? – The Photon Nov 8 '16 at 23:48 • @IgnacioVazquez-Abrams That's kinda what I've already been doing, if I had a decent way to generate really fast rise time pulses, I'd use a diode bridge sampler with some of MACOM's SMS7621-060 diodes seeing as they're$1.60 and are "Suitable for use above 26 GHz" – Sam Nov 9 '16 at 5:32
• @ThePhoton I had a look at some of those Wavecrest analysers, but I'm not 100% sure what they do... is it some kind of jitter analyser? I'm aiming for an order of magnitude higher bit rate (but for a sampled and not real time signal) – Sam Nov 9 '16 at 5:34

Repetitive waveform samplers work well (for repetitive waveforms!)

The comparators specify about 10pS delay variation with input overdrive voltage varying between 50mV and 1v, so if you can tolerate that dispersion on your timings (and it's comparable with the TMU resolution) then you've got a go-er.

I'm building a 32 bit DAC at the moment. Not because I need a 32 bit DAC, but because I want to find out what the real hardware limitations are that will prevent me achieving the theoretical numerical performance. I suggest you build your digitiser in the same frame of mind!

• Just out of curiosity, how are you going to analyze your DAC errors? Particularly, how do you propose to check linearity over the full range? – WhatRoughBeast Nov 9 '16 at 4:28
• I'd be interested in how your 32bit DAC goes as well, it's make a good companion to TI's 32bit ADC (they get ~27bit ENOB) – Sam Nov 9 '16 at 5:35
• @WhatRoughBeast It's obviously got to be sigma delta based. Missing codes, build several with different modulii and compare them, and compare against a big RC. Full range linearity, use a self calibrating ring of resistors that can be driven and sensed at all nodes, as a high linearity very low precision DAC, and integrating the results of a flying reference set to various offset voltages. At the moment, self-consistency between several different methods is all I'm aiming for. The main work has been a number sequence that mitigates tpHL and tpLH skew and charge injection contributions. – Neil_UK Nov 9 '16 at 9:03

I recon it will not work because you will not be able to get the required repeatability. The main problem is the delay of the comparators. They are dependent on the input voltage, and will also vary after each cycle. They will also have process variations.

Also, you are assuming that the signal is the same on each cycle. The signal will probably vary a little bit. If you run a rising edge 1000000 times you'll get a probability distribution.

This is even before you start running into all then non-linearity problems of the comparators, calibration etc.

• If by "repeatability" (or lack thereof) you're referring to the 10ps of dispersion in the comparators... well that just sets an upper limit on the precision of the device. Neither would I need 1'000'000 samples on a rising edge, trying getting better that 6 or 7 bits is likely pointless, so that's 128 samples per edge max, and as I've got 4 channels to play with, I really only need to sample 32 edges to build up a full picture. This is for looking at the overall shape of waveforms, not for taking precision measurments (I'm trying to replicate a \$13'000+ instrument, there will be compromises) – Sam Nov 9 '16 at 5:41