sheldon
|
Eric,
First, comment we are simulating not measuring performance on a test bench so we are taking advantage of the simulator's "ideality" to make the analysis better.
Some general guidelines for a well-behaved FFT of s/h and ADCs are(See Ken's book for more details):
1) The data record for the FFT should include an integer number of evenly spaced(in time) data points. --> This can be a issue, which is one of the reasons why Spectre uses the Fourier integral
2) The data record should have an integer number of cycles of the input and the sample periods. --> This avoids needing to use window functions., that is, there is no spectral leakage so windowing is not required.
3) The ratio of the input frequency to the sample frequency should be a prime number. --> For ADCs in particular, if the input frequency is a related to the sample frequency, fin=10MHz and fs=100MHz. Then only certain codes get exercised and the spectrum has discrete tones at the harmonic frequencies: 10MHz, 20MHz, .. While the SNR calculation is correct, the SFDR and THD calculations are not. All the quantatization noise ends up in the harmonics.
4) The 10 periods is a WAG. Transient analysis takes a while to settle the startup transient. If you think about PSS, the shooting algorithm eliminates the transient start-up and directly calculates the steady-state. In this case, we allowed 10 periods for the circuit to reach "sinusiodal steady-state". In complex designs, circuits can sometimes require long times to settle startup transients.
5) The 9.8ns is a WAG. Assuming that 5ns is allowed for sampling and 5ns are allowed to settle the hold step, then the hold step should be "settled" at about 4.8ns after the start of the hold period. If you wait to long, then the S/H starts to switch back into sample mode and the result is not valid. Since simulators tend to anticipate transistions, you need to allow a little margin.
6) Given your targets of fin=10MHz and fs=100MHz, and your (anticipated) desire for a "reasonable" simulation time. The selections were: a) 128 points for the FFT, power of 2, with reasonable accuracy. Increasing the number of points doubles the simulation and lowers the noise floor. Increasing number of points in the FFT reduces the "resolution bandwidth" of the bin, i.e., less rt Hz, less bandwidth less noise b) Your target input frequency is 10MHz so 13 is a prime number that results in an input tone that is close to your target.
Hope this helps!
Best Regards,
Sheldon
|