Students often ask me, "What classes should I take?" or "What subjects should I study?"
My answer comes in several parts. First, if you want to go into engineering, take some math. Take more than you think you will need. Next, take some accounting. You will need that to figure out where all the money goes in your first startup :-)
Then get a mentor.
This last step is the hardest, but the most important. Start by looking at people you know well. Then expand your scope to teachers, friends, and distant acquaintances. You are looking for people who truly LOVE THEIR JOB.
Few people in the world find a career they love. Those lucky few are hardy, adventurous, risk-taking people. They master a craft, and then, (this is the good part) they pass that knowledge to others.
One such person is Bob Pease, the famous analog guru at National Semiconductor®. Bob's work at National, his magazine columns, and outstanding book "Troubleshooting Analog Circuits" inspire me greatly. When I admire someone's work, I try to find out what makes him or her tick.
In 1996, I met Mr. Pease for lunch in the National cafeteria. Bob greeted me and led the way to a quiet table by the window. We discussed transistor physics, electro-optical conversion, and the increasing use of simulators for circuit design. Then he stopped, looked directly in my eyes, and explained why I should write regular columns for Electronic Design magazine. Beyond the general responsibility shared by all engineers to pass along what they know, he said, "Writing a regular column teaches you something new every month, and that is a good thing." With his encouragement, I did just that. My early writing assignments led to my position today as EDN columnist for Signal Integrity, and finally to this newsletter you are reading today.
There you go. One meeting--one word of advice--changes a whole career. Thank you, Bob.
Analog to Digital Conversion Parameters
I recently taped a session of the BOB PEASE REALITY TV SHOW. Part One of this show will appear on the web Feb. 28. Part Two follows about a month later (see the announcement at the end of this letter). The shows will be archived. You can watch them any time.
On the show, Bob and I discuss a simple A/D application running at 1500 MHz (yes, that's 1.5 Giga-samples per second). The A/D posts its sampled data to a Xilinx® Virtex-4™ FPGA for later analysis. Our discussion touches on the architecture of the demonstration platform, the various amplifier stages used, the performance of the A/D converter, the LVDS bus between it and the Xilinx FPGA, and the USB link over to a personal computer for data analysis.
For the show, we taped a solid hour of round-table discussions plus time in the lab looking at the hardware with Ian King, the board designer. Even with all that, I feel we barely scratched the surface. Mixed-signal applications can be quite complex. I hope to return to Bob's show in the future to discuss other topics.
During our session we used the following five terms, but did not give detailed definitions. I would like to provide those definitions here, along with some hints about how "specsmanship" can distort the results.
SNR signal to noise ratio
This is the ratio of signal power to noise power in a signal (analog or digital), expressed in decibels (dB). It is usually measured using a pure sine wave source. SNR is the fuzziest of the five terms defined here. For example, some people count only "white" noise (specifically, noise uncorrelated with the signal), in the "noise" part of the ratio. Others say signal distortion (harmonics) should be counted as well. When reading an SNR specification make sure you know what effects count as "noise".
Make sure you know the bandwidth over which the "noise" power is defined. For example, suppose you have a 1-watt sine wave carrier at 100 MHz embedded in white noise having a power spectral density of 1 micro-watt/Hz. If you integrate the noise over a bandwidth from DC to 100 MHz, you get a noise power of 100 watts, yielding a negative SNR of -40 dB. If, on the other hand, you plan to band pass filter your signal to a limit of +/-50 KHz on either side of the carrier then you care only about noise in a band that is 100 KHz wide. The noise measured over that much tighter band drops considerably, yielding a positive SNR. Make sure you know over what bandwidth the noise is integrated in any SNR specification.
THD total harmonic distortion
For a pure sine wave excitation, this is that ratio of signal power at the fundamental to the signal power at all harmonics (and, in the case of sampled-data systems, to all aliased versions of the harmonics). Before making this measurement on a system, check the harmonic content of the source. Your system cannot improve on a source with poor THD performance. If you need a better source, try placing a linear band pass filter in series with the source to knock down its harmonics. The THD measurement procedure usually only measures harmonic power in the first few harmonics (4 to 7 typically). When comparing THD figures, check how many harmonics the datasheet numbers cover.
When dealing with sampled-data systems, the harmonics will often be aliased to other, weird frequencies. Sometimes they pop up in places that the test routing does not check. That is a cheap way to fake a better THD number on a specification sheet. Ask if aliased harmonics are counted in the THD number.
SINAD signal to noise and distortion
SINAD is very much like SNR, except that here we are clear that the definition of "noise" includes both white noise and distortion (harmonics). Always ask over what bandwidth the noise is integrated.
What if the signal includes spurious sinusoidal peaks from other, non-correlated sources (like crosstalk from your clock)? Does that count as noise, or distortion, or neither? I do not know. Check your specification sheet.
SFDR spurious free dynamic range
Inject a pure sine wave into your system. The output displays a number of peaks. SFDR measures the ratio of the power in your fundamental to power in the largest (non-fundamental) peak. When calculating SFDR, some datasheets count all the peaks, while some count only peaks not related as harmonics of the data signal. Some count only peaks over certain specific bandwidths (that might be the case when looking at a mixer). Make sure you know which peaks are reported in this number.
ENOB effective number of bits
This is the measurement digital people like best. It condenses everything down into one, easy-to-grasp, simple, wrong number. I say it is wrong because it is so non-specific about what factors are counted. In concept, you first measure one of the other definitions (typically SINAD) and then figure out how quantization levels you would need in an "ideal" A/D converter to attain that same level of performance. The ENOB is the log-base-two of the number of quantization levels required. A good concept, but useful only if two competing manufacturers base their ENOB calculations on the same type of measurement.
In all cases, the measurements may be taken once, or several times and the results averaged. On the show, we used a measurement package called "WaveVision" to do the calculations.
I hope you enjoy watching these sessions as much as I enjoyed making them.
Dr. Howard Johnson
National Semiconductor is a registered trademark of National Semiconductor Corporation. Xilinx is a registered trademark, and Virtex-4 a trademark, of Xilinx, Inc. All other brand or product names are trademarks or registered trademarks of their respective holders.