1/15/2025

EECS 6611

Mixed-Signal Microsystem Design

Chapter 4: Dynamic Performance Metrics

Time vs Frequency Domain

• Time Domain
• We look at transient waveforms and how they settle
• Glitch impulse, aperture uncertainty, settling time, …
• We'll look at these later, in the context of specific circuits

• Frequency Domain
• We apply one or more sine wave tones and look at the spectrum
• Spectral performance metrics
• Important to realize that both static (DNL, INL) and dynamic errors contribute to frequency domain non-
ideality
• Nonlinearity introduce new frequencies ➔ the cause does not have to be frequency-related.
• We have a black-box approach here.

1
 1/15/2025

Spectral Metrics
• SNR - Signal-to-noise ratio

• SNDR (SINAD) - Signal-to-(noise+distortion) ratio

• ENOB - Effective number of bits

• DR - Dynamic range

• SFDR - Spurious free dynamic range

• THD - Total harmonic distortion

• ERB - Effective Resolution Bandwidth

• IMD - Intermodulation distortion

• MTPR - Multi-tone power ratio

• … why so many?

DAC Tone Test/Simulation Setup

2
 1/15/2025

Typical DAC Output Spectrum

• Real measurement from lab bench

• No reconstruction filter ➔ this is the staircase waveform.
• Hence, the images.
• How do we put a number for the quality of this converter?
Hendriks, "Specifying Communications DACs, IEEE Spectrum, July 1997] 5

ADC Tone Test/Simulation

Or these (e.g., from tte)

SR DS360

10 mHz to 200 kHz range

<-100 dBc distortion (to 20 kHz)
20 µVpp to 40 Vpp output range
25 ppm frequency accuracy This is a sequence of
numbers➔ can’t be
measured with lab
• This is oversimplified. instrument.

• What is the requirement for the sine wave generator?

• What else?
6

3
 1/15/2025

Discrete Fourier Transform Basics

• DFT takes a block of N time domain samples (spaced Ts=1/fs) and yields a set of N
frequency bins

• Bin k represents frequency content at k·fs/N [Hz]

• DFT frequency resolution
• Proportional to 1/(N·Ts) in [Hz/bin]
• N·Ts is total time spent gathering samples
• If I want high resolution DFT ➔ More and more samples

• A DFT with N=2integer can be found using a computationally efficient algorithm

• FFT = Fast Fourier Transform

MATLAB Example
clear;
N = 100;
fs = 1000;
fx = 100;
x = cos(2*pi*fx/fs*[0:N-1]);
s = abs(fft(x));
plot(s, 'linewidth', 2);

• MATLAB gives us a two-sided spectrum → we only

want to deal with real signal (no complex signal)
• Amplitude needs to normalized.
• Y-axis in log scale.

4
 1/15/2025

Normalized Plot with Frequency Axis

N = 100;
fs = 1000;
fx = 100;
A = 1;
x = A*cos(2*pi*fx/fs*[0:N-1]);
s = abs(fft(x));
%remove redundant half of spectrum
s = s(1:end/2);
%normalize magnitudes to dBFS
s = 20*log10(s/A/N*2);
%frequency vector
f = [0:N/2-1]/N;
plot(f, s, 'linewidth', 2);
xlabel('Frequency [f/fs]')
ylabel('DFT Magnitude [dBFS]') - What are the additional stuff?
9

Interesting Example
• Change fx to 101.
• What happened?

10

10

5
 1/15/2025

Spectral Leakage
• DFT implicitly assumes that data repeats every N samples

• A sequence that contains a non-integer number of sine

wave cycles has discontinuities in its periodic repetition
• Discontinuity looks like a high frequency signal component

• Power spreads across spectrum

• Two ways to deal with this

• Ensure integer number of periods in your simulation.

• Windowing

11

11

Solution 1
N = 100;
cycles = 9;
fs = 1000;
fx = fs*cycles/N;

• Usable test frequencies are limited to a multiple of

fs/N.
• Bin k represents frequency contents of k/N → This
makes the frequency on top of the bin.

12

12

6
 1/15/2025

Solution 2
• Spectral leakage can be attenuated by windowing the time samples prior to the DFT

• Windows taper smoothly down to zero at the beginning and the end of the observation
window
• More weight to the data points at the center of sample space.

• Time domain samples are multiplied by window coefficients on a sample-by-sample basis

• Means convolution in frequency
• Sine wave tone and other spectral components smear out over several bins

• Lots of window functions to chose from

• Trade-off: attenuation versus smearing

• Example: Hann Window

13

13

Hann Window
N=64;
wvtool(hann(N))

14

14

7
 1/15/2025

Hann Window Result

N = 100;
fs = 1000;
fx = 101;
A = 1;
x = A*cos(2*pi*fx/fs*[0:N-1]);
s = abs(fft(x));
x1 = x.*hann(N).';
s1 = abs(fft(x1));

• Better but not the best.

• Tone energy is split out over 3 bins (because of Hann)
• Sharper shape can be achieved with lots of more samples.
• So why windowing?

15

15

Integer Cycles versus Windowing

• Integer number of cycles
• Test signal falls into single DFT bin
• Requires careful choice of signal frequency
• Ideal for simulations
• In lab measurements, can lock sampling and signal frequency generators (PLL)
𝐽
• "Coherent sampling“: 𝑓𝑖𝑛 = 𝑀 𝑓𝑠 , for M samples and GCD(J,M)=1

• Windowing
• No restrictions on signal frequency
• Signal and harmonics distributed over several DFT bins
• Beware of smeared out nonidealities…
• This could result in harmonics that look smaller but are in fact smeared out.

• Requires more samples for given accuracy

• More info
• http://www.maxim-ic.com/appnotes.cfm/appnote_number/1040

16

16

8
 1/15/2025

Example

• Now that we’ve "calibrated" our test system, let's N = 1024;

look at some spectra that involve nonidealities. cycles = 17;
fs = 1000;
fx = fs*cycles/N;
• Where is my ADC in this code? LSB = 2/2^10;
• What is the resolution?
%generate signal, quantize and take FFT
x = cos(2*pi*fx/fs*[0:N-1]);
• First look at quantization noise introduced by an x = round(x/LSB)*LSB;
ideal quantizer. s = abs(fft(x));
s = s(1:end/2)/N*2;

% calculate SNR
sigbin = 1 + cycles;
noise = [s(1:sigbin-1), s(sigbin+1:end)];
snr = 10*log10( s(sigbin)^2/sum(noise.^2) );

17

17

Spectrum with Q Noise

• Spectrum looks fairly uniform

• Signal-to-quantization noise ratio is given by power

in signal bin, divided by sum of all noise bins

• Expecting
SQNR= 10*6.02+1.76 = 61.96dB

• Noise floor of spectrum is around -80dBFS

• Why not -62dB?

18

18

9
 1/15/2025

Spectrum with Q Noise

• Spectrum looks fairly uniform

• Signal-to-quantization noise ratio is given by power

in signal bin, divided by sum of all noise bins
Why these peaks
and troughs?
• Expecting
SQNR= 10*6.02+1.76 = 61.96dB

• Noise floor of spectrum is around -80dBFS

• Why not -62dB?

19

19

FFT Processing Gain

20

20

10
 1/15/2025

DFT Plot Annotation

• DFT plots are fairly meaningless unless you clearly specify the underlying conditions

• Most common annotation

• Specify how many DFT points were used (N)

• Less common options

• Shift DFT noise floor by 10log10(N/2)dB
• Normalize with respect to bin width in Hz and express noise as power spectral density
• "Noise power in 1 Hz bandwidth"

21

21

Periodic Q Noise
• Same as before, but cycles = 64 (instead of 17)

• fx = fs⋅64/2048 = fs/32

• Quantization noise is highly deterministic and periodic

• For more random and "white“ quantization noise, it is

best to make N and cycles mutually prime
• GCD(N,cycles)=1

22

22

11
 1/15/2025

More Realistic Spectrum

• Fairly uniform noise floor due to additional electronic noise

• Harmonics due to nonlinearities

Why?

• For SNR: total noise power includes all bins except DC,
signal, and 2nd through 7th harmonic
• Both quantization noise and electronic noise affect SNR

23

23

SNDR and ENOB

• Noise and distortion power includes all bins except

DC and signal.
• Why not DC?

24

24

12
 1/15/2025

Effective Number of Bits

• Is a 10-Bit converter with 47.5dB SNDR really a 10-bit converter?

• We get ideal ENOB only for zero electronic noise, perfect transfer function with zero INL, ...

• Low electronic noise is costly

• Cutting thermal noise down by 2x, can cost 4x in power dissipation (EECS 6613)

• Rule of thumb for good power efficiency: ENOB < B-1

• B is the "number of wires" coming out of the ADC or the so called "stated resolution"

25

25

ENOB Survey

R. H. Walden, "Analog-to-digital converter survey

and analysis," IEEE J. on Selected Areas in
Communications, pp. 539-50, April 1999

Does it mean that the last 1 or 2 bits are random?

26

26

13
 1/15/2025

Dynamic Range
• Range of signals that you can resolve with SNR=0;

• DR:
• From: the point that signal and noise power are equal
(i.e., SNR=0dB)
• To: the point you start to decrease in terms of SNR.

• Ideally, peak SNR=DR but because of the saturation

effect → DR tends to be 1-2 dB larger.
• In DelSig the overloading is gradual ➔ people tend to
report DR just to look better.

• For non-uniform sampling, the increase might be

nonlinear → peak SNR could be much smaller
thant DR

27

27

SFDR

• Largest spur is often (but not necessarily) a harmonic of the input tone
• Important metric for radio receivers.

28

28

14
 1/15/2025

THD

• Signal to distortion ratio:

• Total harmonic distortion:

• By convention, total distortion power consists of 2nd through 7th

harmonic.

• Where are the 6th and 7th harmonics here?

29

29

THD

• Signal to distortion ratio:

• Total harmonic distortion:

• By convention, total distortion power consists of 2nd through 7th

harmonic.

• Where are the 6th and 7th harmonics here?

30

30

15
 1/15/2025

Harmonic Aliasing
• If the main tone is at a higher frequency
• Harmonics can appear at "arbitrary" frequencies due to
aliasing.

f1 = fx = 0.3125 fs
f2 = 2 f1 = 0.6250 fs → 0.3750 fs
f3 = 3 f1 = 0.9375 fs → 0.0625 fs
f4 = 4 f1 = 1.2500 fs → 0.2500 fs
f5 = 5 f1 = 1.5625 fs → 0.4375 fs

Why anti-aliasing filter didn’t remove these tones?

31

31

Intermodulation Distortion

• IMD is important in multi-channel communication systems

• Second order products not so important, why?
• Third order products are generally difficult to filter out.
• Shows how your circuit performs in the presence of an interference.
32

32

16
 1/15/2025

MTPR

• Multi-Tone Power Ratio (very application-specific).

• Very broadband communication systems many many different tones (e.g. OFDM, DSL).
• Apply many tones at different frequencies while leaving out two tones.
33

33

Frequency Dependence

• All of the above discussed metrics generally depend on frequency

• Sampling frequency and input frequency
• Why single ended is much worse?
• Why the trend is not smooth and goes up and down?

34

34

17
 1/15/2025

ERBW
• A single number for a very complicated concept.

• Defined as the input frequency at which the SNDR of a converter has dropped by 3dB
• Equivalent to a 0.5-bit loss in ENOB

• ERBW > fs/2 is not uncommon

• especially in converters designed for sub-sampling applications

35

35

Real Example (TI ADS5541)

• You can tell input frequency is more effective that sampling frequency.
36

36

18
 1/15/2025

Real Example (TI ADS5541)

• Same general trend but much more random stuff.

37

37

Relationship Between INL and SFDR

• The INL of an ADC often takes on a quadratic or cubic “bow”
• Meaning that the transfer function can be approximated by x+a2x2
or x+a3x3

• The resulting harmonic distortion usually sets the low frequency

SFDR

• The expression below provides an analytical relationship

between the peak INL due a cubic bow and the resulting HD3

• Rule of thumb: SFDR ≅ 20log(2B/INL)

• E.g. 1 LSB INL, 10 bits → SFDR ≅ 60dB

38

38

19
 1/15/2025

SNR Degradation due to DNL “Noise”

• Another reason why 10 wires won’t result in 10 real bits.

• Your Q-error with DNL is no longer bounded by ±0.5 LSB.

• Intuitively, lower probability as we get to ±1 LSB

• Convolving two squares (ideal quantizer and uniformly-distributed

DNL)

• Lesson: If you have a really bad DNL

• e.g., non-zero DNL across many codes

➔ you should expect a very bad ENOB → another sanity

check.

39

39

20