I build a PCB with the DAC3174 from TI and I'm using it initially to test it as a channel, like shown in the diagram from the datasheet:

enter image description here

The analog signal is mixed with the PLL carrier frequency (at 2.4GHz) and I'm connected to the modulator (mixer) output via SMA connector to the R&S®FSW Signal- und Spektrumanalysator.

For testing it, I'm generating in MATLAB random digital data and applying 16-QAM modulation. The signal is then upsampled (factor 4) and the root-raised cosine filter is applied. The digital data is then mapped to the DAC's resolution and transmitted. MATLAB code is from both these examples (here & here)

Receiver: DAC sampling rate is 100MHz and the oversampling factor is 4 (Symbol rate is therefore 25MHz). The first measurement showed EVM of 82%:

With the real part of the transmitted signal:

I set the filter length from 40 to 17, since the transient response was too long here. The rolloff factor (Beta) is 0.25. I got the following response with EVM of 25%:

Of course setting beta to a smaller value would improve the EVM (0.1 to EVM of 15%). But the middle point of the 16-QAM is messing up the evaluation of the EVM in the FSW. I actually should get something around 5-7% What is the cause of the middle point here exactly and how would I get rid of it? I'm thinking of replicating the signal twice on the left and right sides and only filtering the middle part of the signal.


1 Answer 1


It looks like inevitable filter transient response is the source of the zeros. You see this in the time domain waveform shown where all the initial samples are zeros. The filters initial state is zero, and these are those values initially comping out of the filter as the data moves through the filter. I also see zeros at the end of the time domain waveform which I assume is just the end of a shorter transmission. Ensuring those samples are not included in the EVM computation should resolve the issue; based on the plots showing the constellation I can see that the EVM should be much better (<3% I am guessing).

This theory is also consistent with the EVM being better when the filter length was reduced, as that would shorten the length of the initial zeros accordingly.

(It wasn't mentioned, but there is an additional RRC filter in the receiver, as it would be the cascade of two RRC filters that would result in the zero ISI result-- if this wasn't happening here, the constellation shown would appear much worst).

  • $\begingroup$ Hey! Thank you for the answer :) Yes, you're right. I set everything the same on the receiver (FSW) with the RRC filter, I forgot to mention that sorry. You are also correct about the zeros at the beginning. I'm searching for a workaround just to show the real EVM, that's why I wanted to take these zeros and put my signal there again once on the left and once on the right side, so the filter convolution would go through the signal instead of the zeros. But I'm not sure, how to exactly get that. Do you know what I mean? $\endgroup$ Apr 24 at 13:30
  • $\begingroup$ You just truncate the result so those zeros are not included. $\endgroup$ Apr 25 at 12:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.