ADC sampling problem using FPGA

Question

I am trying to sample sinusoid signals using AD7928, which is a 1 MSPS ADC. It is connected to an Altera Cyclone V FPGA, and runs at 20 MHz master clock. Data rate is 1.25 MHz and that's due to the fact that the ADC outputs a new sample every 16 cycles. Using SignalTap Logic Analyser, I monitor the output of the ADC, and it is clocked with the same data rate frequency (1.25 MHz) to meet Nyquist criteria. Input is a 440 Khz sine signal (under Nyquist's frequency 500 KHz), however, it looks distorted in SignalTap, figure below depicts it. 440 KHz Sine

If I apply a higher frequency signal like 1019 KHz, it looks like a regular sinusoid signal. 1019 KHz Sine

As I change the frequency, signal dilates and contracts.

What do you think is the reason?

• Grabul: Is it really distorted ? At 440KHz for 1Msps, you have barely more than 2 samples per period. The beat frequency (larger ondulations) is around 11 samples, so input frequency should be around 1MHz/2 * (10/11)=454kHz. The lower signal is aliased.
• vicatcu: don't you have to sample at least twice the maximum spectral frequency of your siganl to meet Nyquist criteria?
• Chris Stratton: While it's not the source of your problem (which as has been pointed out is merely one of interpretation), it sounds like you are overclocking your adc by 25% (1.25 MSPS vs 1 MSPS) - it may or may not give good results, and may vary from chip to chip.
• Siraj Muhammad: @TEMLIB I didn't get your point about "beat frequency" and what is the 454 KHz signal? would you please elaborate?
• Grabul: @Siraj : Imagine tuning a guitar, with two strings playing the same note. If the two strings are not perfectly tuned, you will hear a beating modulation whose frequency is the difference between the two strings frequency. Here, we see a similar modulation between the 440kHz sine and the 1Msps/=500kHz sampling frequency. The envelope period is about 11 samples, there is a 1/11 frequency difference.

Answer 1

The problem is that SignalTap is not applying the correct reconstruction algorithm when displaying your waveform. Instead of using a brick-wall low-pass interpolation filter, it's simply using a zero-order hold, which turns each sample into a horizontal line segment.

The sample points themselves really are proper samples of your 440 kHz sinewave. You can prove this to yourself by working it out on paper: draw several cycles of a sinewave that has a period of 2.27 µs, and then show where samples taken at 0.8 µs intervals would fall. Draw a horizontal line from each sample to where the next sample is taken, and you should end up with something that looks very much like what SignalTap is showing you.

• Siraj Muhammad: Thanks for this explanation. Actually I figured out that I posted the wrong picture, it was for a signal captured in SignalTap using a 20 MHz clock for SignalTap, i.e. I used a higher clock frequency than data rate frequency. Now I changed it, however the same result. I think that won't change anything in your explanation, right?
• Dave Tweed: Right. Using the higher-frequency clock just made the "horzontal lines" more obvious, since each sample value was repeated 16 times. But the underlying theory remains the same.
• Siraj Muhammad: Actually I am trying to use the ADC to demodulate FM signals by means of a wired channel. Is not there any solution to overcome this problem using this particular ADC? or should I use faster one?
• Dave Tweed: That's a rather complicated topic, and should probably be asked as a completely separate question, and it should be perhaps be asked in DSP.SE, where you'll probably get a better answer. The issue is that the information you want is encoded in the zero-crossings of the signal, and the raw ADC sample rate is not giving you a lot of detailed information about the timing of the zero crossings. It takes some signficiant processing to do good FM demodulation that close to the Nyquist bandwidth of the ADC.

Answer 2

Dave Tweed's answer is completely correct (+1) - I'll just add that (for given results) the sophistication of the reconstruction filter is going to need to be a lot less if you can sample significantly above the Nyquist frequency of the highest frequency component in the signal.