PGA460: Frequency offset in high frequency mode

Hi Sergey,

I will need a moment to review this data set and confirm the behavior on our bench setup. The roll-over at 362kHz is unexpected.

In the meanwhile, here is a more detailed explanation on how the digital coefficients are determined:

The Band-Pass filter works based on digital coefficients B1, A2 and A3. These are calculated based on the burst frequency (FREQ parameter) and Band-Pass bandwidth (BPF_BW parameter). The A2 coefficient is calculated as following:

• A2= (0.6036*cf^2) +( 2.7401*cf) -32615           ….           For BPF Bandwidth  2kHz.
• A2= (0.5999*cf^2) + (2.7217*cf) -32415           ….           For BPF Bandwidth  4kHz.
• A2= (0.5965*cf^2) + (2.6791*cf) -32217           ….           For BPF Bandwidth  6kHz.
• A2= (0.5928*cf^2) - (2.6726*cf) -32024            ….           For BPF Bandwidth  8kHz.

Where “cf” is the filter center frequency in kHz (Defined by the FREQ parameter).

The B1 and A3 coefficients are calculated purely based on the filter band-width:

• B1=199.7*bw+12
• A3=-399.3*bw+65512

Where “bw” is the filter band-width in kHz (Defined by the BPF_BW parameter). Since there will be only four B1 and four A3 coefficients it might be worth to simply store the values instead of doing the calculation. In this case, below are listed the correct values for the B1 and A3 coefficients:

• B1 = 409 ; A3 = 64718        ….           For BPF Bandwidth 2kHz.
• B1 = 813 ; A3 = 63909        ….           For BPF Bandwidth 4kHz.
• B1 = 1213 ; A3 = 63111       ….           For BPF Bandwidth 6kHz.
• B1 = 1607 ; A3 = 62322      ….           For BPF Bandwidth 8kHz.

It is important to note that these are scaled-up coefficient values. The true B1, A2 and A3 values are always low number values (typically < +/- 2). The scaling is the following:

• B1 = B1_real * 2^16
• A2 = A2_real * 2^14
• A3 = A3_real * 2^16

In the digital filter implementation the values are scaled-down right before their use. IMPORTANT: The above coefficient formulas are only applicable for a sample rate of 1us.

Hello Akeem,

As an interim solution, we did our own polynomial fit on the simulated data, made possible by the scaling coefficients you released previously. Coefficients are below:

A2 is calculated as a third degree polynomial: a2 = a * fc^3 + b * fc^2 + c * fc + d, with fc - center frequency in hertz (not kilohertz as with TI polynomial).

We were able to achieve an acceptable fit across the most of the frequency band, however, there are some issues still, for instance, there is significant gain peaking-like behaviour below 100 kHz and above 420 kHz, which manifests as elevated noise floor and, consequently, degraded SNR. While this solution is better than our previous method of tweaking a2 by hand, we would prefer a solution with better behavior on 450 kHz.

Hi, Sergey!

Hi, Akeem!

I submit the behavior of the chip that Sergey described...

Sergey! Thanks a lot! For two monthes I can't understand what I'm doing wrong... After reading your post I try to change A2 coeff. and find out that on 250.8 kHz max response is at A2_ti-6900=0xFDDC. (That was measured before you post your fomular. By the way: Your coeff. for that frequency is 0xfffc  and response with it is 2dB less than with 0cfddc. ) I'm planning to use your formular, but i'll try to find why is that difference in responce...

We would like to repeat that we are still waiting for TI to respond regarding this issue.

As I have stated before, we have three frequencies that we are interested in:

1. 60 kHz. PGA460 meets our requirements at 60 kHz as is.
2. 250 kHz. Currently we have a stopgap measure implemented, as I've detailed in my previous post. As Kostya correclty notes above, this measure is not perfect. We need to have this issue resolved properly by TI before we go further towards production.
3. 455 kHz. Performance at 455 kHz does not meet our requirements - or even PGA460 specifications - with or without or modification.

We would like to see TI either confirm that this is indeed a deviation from the stated specification, and what timeframe is to be expected for a workaround, or confirm that this behavior is intended. We have a tight deadline on this project, and I would like to know whether we need to redesign with a different AFE as soon as practical.

Hi Akeem,

Thank you for the progress report.

Attached is the worksheet I've used for polynomial fitting. It may be of use to you, as it contains the simulated data: for every defined combination of B2 and A3 and for every possible value of A2 I've done a frequency sweep, analyzed the response, and recorded frequency and amplitude of its peak, if it exists. It is less noisy, more accurate, has more data points, and these data points are distributed evenly across entire range, without compression towards high frequencies displayed in the spectrum analyzer data in my initial post. Do note that A2 values are signed, because I've found that fitting in two's complement directly produces bad fits due to discontinuity around the zero, and all frequencies are in hertz.

As soon as you have uploaded the new GUI and tables, I'll recheck them on our bench and report back.

worksheet.xlsx

Hi Akeem,

Your new polynomial works perfectly: we observe a SNR improvement of approximately 3 dB at 250 kHz, and the performance at 450 kHz is now satisfactory.

Thank you.