[time-nuts] On low-voltage TAC/TDCs for a GPSDO
bruce.griffiths at xtra.co.nz
Thu Aug 12 23:37:11 UTC 2010
1) Use a 74AHC05 for Q1 and Q2.
2) Switch the current source at the emitter node and only turn on the
current source when charging the capacitor.
This will increase the available TAC output voltage range and/or improve
the linearity by eliminating the diode.
However the capacitor discharge switch should be turned off before
charging the capacitor.
A stable fixed delay of a few (10ns??) before switching on the current
source is required.
3) Replace the current source with a resistor.
The resultant nonlinearity is well defined and software correction
should be relatively easy.
4) If the ADC(s) have a sufficiently wide full power bandwidth then one
could just sample a pair of quadrature phased 250kHz sinewaves.
Extend the range by sampling (synchronise the input sampling edge to the
counter clock first) a counter clocked at 250KHz.
Initiate the sampling with the signal edge to be time stamped.
If the GPSDO is used to clock the microprocessor, counters and produce
the quadrature sinewave outputs then only a single TDC (time to digital
converter) is required.
Measuring negative time intervals should not be necessary as the TAC (or
other TDC) should be used merely to measure the delay of a synchroniser
the output of which is used to synchronously sample a counter clocked
with the same clock as the synchroniser.
J.D. Bakker wrote:
> Hello all,
> I'm working on Yet Another DIY GPSDO, and one of the issues I've been
> looking into is a TAC/TDC to do sawtooth correction on the measurement
> of the GPS PPS signal. I'd like to stick with a 3.3V supply for most
> of the circuit, and several of the TAC designs that have been
> discussed here in the past run into trouble at such low voltages
> (mostly through VBE drops).
> To start with the context: I'm planning to use a microcontroller with
> a built-in dual 12-bit 2MSPS ADC. I'd like to not use anything that's
> not available at Digi-Key or Mouser, and keep the SMD pitch >=0.8mm
> (with a possible exception for dual transistors in SOT-23-6). That way
> the design shouldn't be too hard for others to replicate.
> I'm aiming for a TAC accuracy of 1ns, allowing for one or a few
> calibrations between PPS pulses. Minimum full-scale range should be
> +/- a few hundred ns, to allow for outliers. (The plan is to have an
> initial FLL for coarse locking, and have the PLL kick in after that).
> I'm penciling in an ADC reference voltage of 2V, as that's commonly
> available and leaves enough headroom to use the current sources in
> their most linear range.
> I've attached a diagram that reflects a few of my current thoughts.
> - Circuit 1 is the traditional TAC. Before the start of the cycle Q2
> conducts, discharging C1 and shunting I1's current to ground. At this
> point the ADC can measure the voltage drop across C1/Q2 to eliminate
> that offset. Taking nSTART low puts Q2 into high-impedance, and I1
> charges C1 through D1 until STOP is raised causing Q1 to shunt I1's
> current to ground. At this point the ADC samples the voltage across
> C1, which is proportional to the time between START and STOP (modulo
> offset and nonlinearities).
> This circuit is well known to work (although it is more common to use
> Q1 for both START and STOP and to limit Q2 to ramp discharge duties).
> Downsides are that negative time offsets cannot be measured directly,
> and the constant output voltage offers little room for increased
> precision through sample averaging, unless the ADC's input noise is
> large compared to its LSB size. For the same reason there is no easy
> way to reduce the effects of ADC INL/DNL.
> - Circuit 2 works in a similar way, except that the ramp isn't
> terminated by a STOP signal but is allowed to run freely until I2
> saturates. The ADC is set to sample continuously, taking multiple
> samples of the ramp, and the microcontroller interpolates the
> resulting values to determine the elapsed time between an internal
> time reference point and the START signal.
> This circuit is fairly simple, and has the advantage that there is no
> hard limit to its range. Curve-fitting the sampled values increases
> precision and reduces the effects of INL/DNL. On the other hand, ADC
> aperture jitter and offset have a direct impact on resolution.
> - Circuit 3 expands on this approach by having dual ramp generators,
> and having the ADC measure the voltage difference between the two.
Not a good idea, as this requires accurate matching of the gains of the
Its better to sample each TAC output individually as this allows
software correction for gain mismatch (and nonlinearity) before subtraction.
Software correction is better than using trimpots or similar as the
parasitics etc associated with trimpots are eliminated.
> This approach is the only one of the three that can directly measure
> negative time offsets, allowing a regenerated pulse to be directly
> compared with the GPS' PPS. A small difference in ramp rates,
> unavoidable in practice, actually helps to average out DNL and is
> easily corrected in calibration. Sampling time uncertanties have less
> impact than in Circuit 2. Then again, it may be difficult to reliably
> detect the start/end-of-ramp points from the samples alone. Total
> range is relatively limited, and due to the differential measurements
> it is harder to reduce current source nonlinearities in software.
> Any thoughts? At this point I'm tempted to build a hybrid of 2 and 3,
> using one of the microcontroller's ADCs in each mode.
> I've not seen prior work on the ramp-approach, although it's a close
> cousin to the centroid pulse timing method
> (<http://www.febo.com/pipermail/time-nuts/2006-September/021765.html>). Has
> anyone seen it before (and possibly shot down due to major
> deficiencies)? It seems too obvious to not have been considered by
It was usually not feasible as the ADC's typically used had insufficient
input power bandwidth.
The settling time and power bandwidth of any buffer amplifier between
the ramp capacitor and the ADC has also to be considered.
If one uses a capacitive input charge redistribution ADC connected
directly to the ramp capacitor then the sampling process itself
transfers charge from the ramp capacitor to the sampling capacitor.
Software compensation for this effect may be required as the transferred
charge depends on the number of samples from the ramp start to the
You will also need to ensure that the current source recovers
sufficiently quickly from saturation.
Another issue is to limit the discharge current flowing in the discharge
Often a 2 step discharge is used.
A switch with a series resistor is used to discharge the capacitor to
the point at which the second switch can be turned on to complete the
discharge without excessive curent flowing in this switch. See the
HP53131A/2A schematics for an example.
> (Notes: These are initial rough sketches. The ramp current has not
> been optimized yet; I have an unsubstantiated feeling that
> brute-forcing it with a higher current and larger cap may well help to
> swamp some of the nonlinearities. I've mostly picked 1V/us ramp speed
> out of the air because it gives me 4-5 samples @2MSPS which is a
> workable number to do curve fitting on. Also not sure whether I'll use
> the simpler one-transistor current source or the hi-Zout mirror with a
> current source derived from the ADC's reference. The FETs may end up
> being implemented as single-gate /OE drivers. I'll do a more complete
> write-up on the entire GPSDO later).
More information about the time-nuts