HSDCPro Data Windowing

Hi Romain,

I do not know the exact Blackman window as we use a DSP library and do not perform the FFT within our GUI code.  I will try to find the exact Blackman that is used, however we do sum the adjacent  bins in the various windows for the total fundamental power.  For the rectangular window there is no summation as the signal should be coherent when using a rectangular window and we do not expect any spectral leakage.

This is described in the HSDC Pro users manual.  A scaling factor is also applied to account for the signal loss due to the windowing.  In essence what is reported is the expected power of the tone accounting for windowing.

Ken.

Hi Ken,

Thanks for your reply.
If I understand well, for the power calculation, I should set to 0 the "number of bins to remove on either side of fundamental". Am I correct? The point is that at 350 MHz (-1 dBFS) the SFDR is highly degraded if I keep all the bins on either side of fundamental. Using Blackman window gives datasheet results.
Also, when I change input frequency from 350 MHz to the coherent frequency calculated by the GUI (i.e. 350.15869141 MHz @4 GSps) the results are confusing:
- 350 MHz (no bin removed): Rectangular window gives -1.04 dBFS (SFDR = 52.27 dBFS) => let say it is the reference.
- 350 MHz (no bin removed and same power on the signal generator): Blackman window gives -1.04 dBFS but the max on the curve is at -3.4 dBFS (SFDR = 70.85 dBFS)
- 350.15869141 MHz (no bin removed and same power on the signal generator): Rectangular window gives -3.5 dBFS with bad FFT (SFDR = 6.92 dBFS on the fundamental side)
- 350.15869141 MHz (no bin removed and same power on the signal generator): Blackman window gives -1.1 dBFS (SFDR = 72.1 dBFS) but the maximum of the curve is at -4 dBFS.

How should I interpret these different results.

Last question: how input signal and clock signal are made coherent if they are?

Thanks again for your help.

Best regards,
Romain

Hi Romain

Since you are using the internal clocking it will not be possible to have a coherent input signal frequency and therefore Rectangular windowing should not be used.

If you want to use Rectangular windowing and a coherent input signal you will need to modify the board to allow external clocking to be used. The instructions for doing this are in the ADC12J4000EVM User's Guide starting on page 21. http://www.ti.com/lit/ug/slau551/slau551.pdf

If you use the other FFT window settings then adjacent bins are integrated (excluded from notching) to determine the total signal power and the fundamental and in the harmonics. This is necessary because those windowing techniques spread the signal energy into multiple bins. If Rectangular windowing and a coherent signal are used, then all of the power of the single tone input will fall into a single FFT bin. Likewise for harmonics of the fundamental.

Please see section A.3 of the High Speed Data Converter Pro GUI User's Guide for details regarding which adjacent bins are integrated.

Section A.2.1 lists the magnitude correction factors applied for the different windowing techniques. The combination of these correction factors, and the spectral spreading caused by the non-Rectangular windowing methods causes the integrated power to differ from the height of the peak shown in the FFT.

This plot shows the effect of scaling for a 607.77MHz input at -1.57dBFS. The peak of the curve appears to be at -4.3dBFS while the integrated power reported is -1.57dBFS. This is the expected and correct behavior.

If you are seeing something other than this, or getting worse performance than expected please let us know.

Is possible please provide an FFT plot of the results you have questions about.

Best regards,

Jim B

Hi Jim,

Thanks for your clear feedback. It helps.

I typically observe what you show on the FFT plot.

One last question: From your screenshot, you don't check "Auto calculation of coherent frequencies". Is it because you use internal PLL so that you don't need to specify it to the GUI?

Best regards,

Romain

Jim,

Please find attached the screenshots I took for 350 MHz and 350.15869141  MHz (calculated coherent input frequency). Even if I can't use coherent signals (I am using internal PLL) I played with different windowings (Rectangle and Blackman) to compare the results. Measurement at done at same input power (Signal Generator -> Pass band filter -> ADC12J4000).

In "Notch frequency bins", I chose to remove 10 bins on either side of fundamental, 0 bin to remove on either side of harmonics, 10 bins to remove after DC, and "Don't integrate".

I think the results are in line with what is expected. What do you think?

Thanks and best regards,

Romain

20150623.zip

Thanks for you feedback and the link to the blog.

Best regards,

Romain