Home
Prox / RFID
Verichips
Ladder Logic
Tube Joints
SolveSpace (3d CAD)
SketchFlat (2d CAD)
Photographs
Miscellany
[trivia] †
Resume / Consulting
Contact Me
Bookmark and Share

Electronic Parts and Construction

Printed Circuit Boards

I typically use two- or four-layer boards. I never autoroute anything. I use half ounce copper unless I am working with high currents, and sometimes even then. (The quickest-turn places won't give you thick copper.) For high-current boards, I draw traces as polygons, instead of as line segments with thickness; this lets me put every available square mil of copper to use.

I expect that I would have to change my habits considerably if I wanted to do complicated digital stuff; as far as I know, that makes heavy use of autorouters, and needs very high layer counts just to fan out the latest BGAs. Everything about that kind of work is expensive and slow: high-end CAD tools, printed circuit boards with three-week lead times, BGA rework, and it just gets worse from there.

I design for 8 mil (thou, whatever; 0.008") traces, with 8 mil clearances between. I try to use an 0.4 mm minimum via, but sometimes must go smaller, to 0.2 mm. I almost always use 0.062" FR-4, unless there is a good reason not to. This gives me wide choice of vendors.

I see no indication that North American printed circuit board vendors are in any way better than Asian. In low production quantities (hundreds), they charge between triple and ten times as much.

I work in EAGLE. Ignoring the autorouter, it seems as good as or better than its competitors. Protel is awful, and ten times the price. PADS doesn't look bad, but I don't see what would justify paying for it.

Suppliers

I'm fond of Digikey. Mouser is the other big small-quantity distributor, but their web site is painful to use. Beyond that, an increasing number of manufacturers have their own web stores for prototype quantities, where you can buy samples with a credit card. The old-fashioned free sample thing would seem to be dying, justifiably; the time to convince them you deserve a free sample costs more than the parts.

Future and Avnet are not horrible to deal with for small (~1000 USD) orders, but more trouble than Digikey or Mouser. I've also used Nu Horizons, with good results. Most large distributors are set up for purchasing departments and net 30 terms; the overhead is not practical unless you're spending thousands of dollars per line item.

McMaster Carr sells hardware and materials (metal foils, adhesives, solvents, plastics, small tools, many things). They are easy to deal with. Other industrial supply companies (MSC, Grainger, etc.) have similar selection and sometimes better pricing, but the paperwork is less convenient.

McMaster Carr and Digikey can both ship pre-cleared to Canada. This means that they collect applicable Canadian taxes, so nothing's due at the border. You pay the taxes either way, but if they're collected at the border, then you will also pay a brokerage fee, typically on the order of fifty dollars. By shipping pre-cleared, you avoid this.

Assembly

Most components can be prototyped with only a soldering iron, tweezers, a loupe, solder, and extra flux. I used to use paste flux, which I scraped on with a sliver of cardboard, but I now tend to use flux pens, which don't make such a terrible mess. I avoid no-clean flux. A microscope makes things go much faster; a suitable instrument will have a long working distance, so that the soldering iron is kept well away from the lens.

Leadless components are harder, even for things that look easy. I mounted some leadless crystals with an iron, and they stopped working two months later. Hot air is necessary. I use the kind of station that blows air from above and below, with nominal thermostatic control; who knows how the air temperature relates to the temperature at the board and chips. (Real multi-zone ovens keep everything close enough to thermal equilibrium that in theory it's all the same temperature. IR reflow ovens run closed loop on a temperature sensor in good thermal contact with a `witness board' of similar construction to the ones being soldered.)

I do not use any placement aids. With care and 10X magnification, it's easy to line 0.5 mm pitch parts up by hand; try pressing your finger on the board, and rolling it towards the chip, to nudge it just a very short distance. Flux is helpful here as well; in addition to its job as flux, the tackiness helps to hold parts in place.

Desoldering braid is essential for cleaning up solder bridges. I like the Chemtronics product; other brands are cheaper, but the flux doesn't seem to work as well. I don't know a good way to desolder large ICs without a hot air station. (On the hot air station, you just melt all the solder at once, and lift the part off.) I've never been able to make the vacuum solder suckers work for me. If I need to clean out a plated through-hole, then I just melt the solder and push a dental pick through.

Zephyrtronics markets low-melting metal wire as a desoldering aid. This used to be bismuth, but they're now shipping indium. The idea is that if you mix the metal in to a solder joint that you want to remove, then the joint will stay liquid, even at fairly low temperatures. This makes it easy to melt all of the solder joints at once, and lift the component off with no damage to the lands. Once the component is removed, you can clean the board with flux and desoldering braid. It's hard to pick up the pure low-melting metal with just copper braid; it melts, but doesn't wet. I've had good luck alloying any leftover metal with additional (ordinary tin/lead) solder, to make it wet the braid. They claim that the residual bismuth/indium does not affect the properties of the reworked solder joint. I'm not sure, but for prototype work it seems fine.

Zephyrtronics sells an assortment of neat products like that, including air-operated syringes (to dispense solder paste or other fluids), and various lint-free swabs and brushes. They also sell hot air stations, and other more expensive equipment. I am not so impressed with those. The quality isn't bad, though a `ZT-1-CLS-DPU Airbath' shot sparks in my face once. (The heating element had burned through; we called them up, and they wouldn't even sell us a replacement part.) It's very expensive, and the Chinese knockoffs that people sell on eBay are a quarter the price.

Commercial assemblers are everywhere, both here and in Asia. Typically you send them design data and (if they don't make them) boards, and possibly a parts kit, if they don't stock them and won't buy them for you. I prefer to make up the parts kit myself, because it gives me a chance to find mistakes. Quality is variable; I've had minor disasters, and witnessed major ones. I am satisfied with yields (before rework) in the high- to mid-nineties, or better on simple boards.

Automated surface-mount assembly, with machine placement and hot air reflow soldering, achieves incredibly good quality. The trouble is usually with the easy stuff, like through-hole connectors that are installed by hand. I also get missing or misaligned components, which presumably were nudged while the operator was moving the panel from the pick-and-place to the reflow oven.

Test Equipment

I like the Tektronix TDS1000 and TDS2000 series; these are modern digital storage oscilloscopes. They are useful up to their rated bandwidth. (This isn't always true; the PC-based oscilloscopes that quote `repetitive bandwidth' are particularly bad.) The triggering is not as flexible as on more expensive models, but can usually be made to work; worst case, it's often possible to modify the software or logic to provide an explicit `trigger' signal on a spare pin.

I almost always use `peak detect' mode; that means that the oscilloscope, instead of plotting a point for each sample in time, corresponding to the voltage measured at some instant in that interval, plots a vertical line, corresponding to the maximum and minimum voltage seen. This eliminates any possibility of aliasing, and shows high-frequency noise even when its period is faster than a single pixel on-screen.

Tektronix also sells digital phosphor oscilloscopes (DPOs), which simulate the characteristics of an analog oscilloscope's CRT. These are more expensive, but tend to come with nicer features, like longer trace buffers. I don't like these, probably because I don't have any intuition for what things are supposed to look like on an analog oscilloscope. I usually get more from a DSO set to `peak detect'.

Hand-held multimeters are basically disposable. You can buy one for less than ten dollars, and it will be accurate to within a few percent. We paid five dollars each for an ugly gray thing, which came with a dead battery and an input resistance of a megohm on the voltage scales. (On the microamp scales, it looks like 1k.) There's an argument for paying twenty dollars, but not thirty.

Resistors

Most people use surface-mount thick-film resistors. Thin-film resistors also exist. The `film' is the resistive element; the difference is not just the thickness of the film, but also the process by which it is made. A `thin film' is vacuum-deposited metal. A `thick film' is ceramic. I've read that thick-film resistors have worse excess noise, but I've never had an application where I could measure a difference.

The most common sizes are 1210, 1206, 0805, 0603, and 0402. (The digits give the size of the resistor; an 0805 resistor is 0.080 by 0.050 inches.) As of this writing, 0603s are probably most popular. 0402s are routine, though. Parts as small as 01005 (0.010" by 0.005") are available, but will increase assembly costs.

I like to use 0603s when I have no reason not to, though at work they standardized on 0805s several years ago. I use smaller components when better electrical performance (more negligible parasitics) is required, especially at microwave frequencies. I use larger components when greater power handling is required, though if I just need to dissipate 1/4 W, then I will often parallel or series two 0805s of twice or half the value.

Different manufacturers quote different power ratings for the same size resistors. I suspect that this reflects differences in the characterization process more than differences between the products.

1%-tolerance resistors are standard, and only slightly more expensive (or even cheaper!) than 5% parts. 0.1% parts are more expensive, but not unreasonable.

Ceramic Capacitors

For small values, most people use surface-mount ceramic capacitors. The size codes are the same as for resistors. These are often referred to as MLCCs, multi-layer ceramic capacitors. Single-layer ceramic capacitors are exotic; they're typically used only at microwave frequencies.

Capacitors are rated by their value (in pF, nF, or uF; though Americans don't like nanofarads, and will typically write 1000 pF, 0.010 uF, 0.1 uF) and by their max rated voltage.

In addition to that, they are rated by their dielectric's properties, using letters like NP0/C0G, X7R, X5R, Y5V, etc. NP0/C0G capacitors have good properties—stable with temperature, very linear, high rated voltages—but are available only in small values, up in to the nanofarads. These capacitors are suitable for use as frequency-selective elements in oscillators, in filters, and in other applications where exact capacitance matters.

X5R and X7R capacitors are less ideal. Their capacitance changes somewhat with temperature, by tens of percent. Their capacitance is also nonlinear; it changes somewhat as a function of the bias voltage across the part. Still, they're significantly cheaper, and available into the tens of uF, though at a very high price. These make good bypass caps, but they are also suitable as real circuit elements (slow filters, etc.) in noncritical applications.

The most exotic dielectrics offer higher initial capacitance, but with many drawbacks; they will quickly lose capacitance with changes in any parameter you can imagine. Y5V capacitors, for example, are specified as losing 90% of their capacitance at half of rated voltage! This means a 10 uF, 6.3 V Y5V bypass capacitor is no better than a 1.0 uF, 6.3 V X5R capacitor, if both are used at 3.3 V. These are good only for bypassing, if that.

Big MLCCs make wonderful output capacitors for switching power supplies; their low ESR greatly reduces ripple and loss. They are a noticeable part of the improved efficiency of prefab DC/DC modules. They're expensive, though, very expensive.

It's hard to find capacitors that are specified at RF. It would be nice to see more plots of impedance versus frequency, but that's not very common. Special `microwave capacitors' come with data, but they're very expensive, and not necessarily that much better than general-purpose parts. I like to measure S12 on a network analyzer, with the capacitor under test shorting the microstrip to ground; the null occurs at the self-resonant frequency.

Tantalums

These are losing favor. They used to be popular for power supply use, where their low ESR made them handy as filter caps in switchmode power supplies and for general bypassing. They're expensive, though. Also, I've read that their use in high-peak-current applications like that is what gave them their reputation for exploding.

Little ones (<10 uF) can probably be replaced with ceramics these days. Big ones can often be replaced with low-ESR aluminum electrolytics; these are much better than they used to be.

Some LDOs will explicitly require a tantalum output capacitor. This is because the tantalum's low-but-nonzero ESR stabilizes the regulator's control loop; either too much (if you used an Al electrolytic) or too little (if you used ceramics) will make it oscillate. You can fake it with an MLCC and a low-ohm resistor, though that seems a little bit silly. You can also buy regulators that don't require this.

Aluminum Electrolytics

These have a terrible reputation, but have improved. I still routinely use large electrolytics for low-frequency bypassing. They are adequate as output capacitors for switching power supplies. Ripple (and capacitor sizing) is determined by ESR and rated RMS current, not by dV = Idt/C. Given ESR, the manufacturer's rated RMS current is implicitly a power rating. When used at full rated current, they will be warm to the touch, but not hot.

An aluminum electrolytic that is rated to handle the output ripple of a typical switching power supply (and especially a flyback converter, or another topology with highly discontinuous output current) will typically be huge enough to make a good transient response easy. (The ESR may also help, to the extent that it adds damping.) I did a software-controlled power supply that regulates fine, with the digital control loop running at only 1500 samples per second.

When capacitor choice is driven by ESR and power dissipation, the large resulting capacitance may start to become a problem, because the inrush current (and stored energy, if you're worried about fault conditions) become very large. I've sometimes used a capacitor with a higher voltage rating than required. In the same case size, this gets lower stored energy, but comparable ESR and rated RMS current.

The above is true for small changes in rated voltage, but not large ones. I'm not sure why; I suspect that it becomes false when the voltage changes enough that the manufacturer uses a different electrolyte. For Nippon Chemi-Con's Alchip MZA line, for example, the `HA0' style package is rated for 0.16 ohms/600 mA RMS, for under 50 V. At 50 V it's 0.34/350, at 63 V it's 0.7/250, and at 80 V it's 1.3/130.

The product IRMS2ESR varies less, but it's not constant. The thermal resistance to ambient should be the same for identical packages, so this means that the allowable temperature rise is not constant. I am not sure if that means that the different electrolytes tolerate different peak temperatures, or whether the RMS current spec is being driven by something other than power.

Rated lifetimes are still bad. It will probably be the first thing that fails. Older families of electrolytics are just awful, and no cheaper. Any family that's rated for ripple current and for ESR/impedance at a high (~100 kHz) frequency is probably fine.

Inductors

In general, these are avoided. Inductors are more expensive and less ideal than either resistors or capacitors. For low-frequency signal processing, an active filter is typically a better idea. It's possible to build a circuit (a `gyrator') that simulates an inductor, in terms of resistors and capacitors and active elements. A textbook method for filter design is to start with an RLC prototype, and replace the inductors with gyrated fake inductors. I don't think that is very common in practice; it would be more typical to design an active filter from the start, using standard biquad sections made from opamps and R's and C's. It's also common to solve these problems using DSP.

For switchmode (or other) power conversion, an active `simulated inductor' would rather miss the point, unless we are happy converting simulated power. Typical switching power supplies use one or more inductors, sometimes as simple two-terminal components, sometimes as multi-winding inductors (or, equivalently, as transformers that store energy). Typical frequencies go from tens of kHz to about a MHz. This means that air-core inductors are not practical; too many turns would be required to achieve a useful inductance.

A ferromagnetic core is required. At very low frequencies, iron is used. This works, but not very well. Iron is ferromagnetic, but it's also an electrical conductor. This means that a time-varying magnetic field induces current in the core, which means that the core burns power as I2R. It's possible to increase the effective resistivity of the core by splitting it into many thin sheets, each electrically insulated from the others. This technique is visible in many 60 Hz transformers and motors. It's useless at typical SMPS frequencies; the sheets would have to be impractically thin.

Up to about a hundred kHz, powdered metal materials are used. These are exactly what they sound like: particles of ferromagnetic but conductive whatever, in a nonconducting but non-ferromagnetic matrix. The particles are small, and isolated from each other. This means that the effective resistivity is high, so the core losses are decreased. The disadvantage is that the effective permeability is also decreased.

At higher frequencies, it's typical to use a ferrite. Ferrites are compounds of ferromagnetic metals that are themselves ferromagnetic, but non-conductive. This is ideal; you get high permeability, so high inductance, but very low core losses. The disadvantage is that ferrites are nonlinear. This means that the permeability of the ferrite decreases sharply as the magnitude of the magnetic field increases. So the inductance will be a function of current; large initially, and then decreasing. This nonlinearity means that an LC filter, for example, would generate harmonics and intermodulation products of its input. Or, in a typical switching power supply, with constant voltage applied across the inductor, the current would increase not as a linear ramp but as a convex function of time, slowly at first and then faster.

The usual way to fix this is to cut a small air gap in the magnetic circuit. This air gap has low permeability (μr = 1, compared to thousands for a typical ferrite); so initially, at low currents, the reluctance of that air gap dominates the total reluctance, even though the air gap is small compared to the total length of the magnetic circuit. Since the permeability of air is very close to constant, independent of the field strength, the inductor will be relatively linear, as desired. At sufficiently high current, the magnetic field strength in the ferrite will get large enough that its permeability will drop sufficiently that the air gap reluctance no longer dominates the total reluctance. At this point, the inductance will begin to drop significantly. A typical manufacturer's datasheet will describe this point as the saturation current, and define it as the current where the small-signal linearized inductance has dropped by some amount from its initial value, typically around 10%. (Of course, the threshold that they choose is arbitrary; but the current corresponding to a ~10% drop is often a reasonable design point for the maximum current through the inductor.)

Typical ferrite cores come in two pieces. The coil is wound on a bobbin; the two halves of the core are then inserted and glued together, closing the magnetic circuit. Custom magnetics (inductors and transformers) are very common. A multi-winding inductor has too many degrees of freedom—turns count for each winding, wire gauge for each winding, core size, core shape, core material, gap width, ...—for an exhaustive catalog of standard parts to be practical.

At high frequency, simulated inductors become impractical, even for signal work. This is because it's not possible to find active components fast enough to simulate them. It might also be necessary to use a real inductor for noise reasons; an active filter adds semiconductor device noise, but a high-Q inductor adds almost none.

Ferrite materials are popular for `signal' inductors at low frequencies; in this application the poor energy storage characteristics are not a concern. At higher frequencies (tens of MHz), air-core inductors are typical. At UHF and higher, microstrip design becomes useful (i.e., the printed circuit board traces are considered as pieces of transmission line, rather than as perfect short circuits). Microstrip inductors become practical at lower frequencies than do microstrip capacitors, for typical printed circuit board geometries and materials.

Air-core chip inductors in small (0402, 0201) packages are also available, with values down to about a nH. These are not necessarily better than a microstrip inductor on low-loss substrate, but they make it easier to change component values without manufacturing a new printed circuit board.

Crystals

Most people seem to use HC49/US through-hole parts. This is the short metal can with two leads 5 mm apart. Tubular packages are popular for the 32.768 kHz `tuning fork' crystals, and also for some higher-frequency parts (that I assume vibrate in the same mode as the ones in flat cans, but just are packaged differently? I don't know, though).

I don't know of a standard for surface-mount crystals. There's one that looks a lot like an HC49/US, but with flat tabs instead of leads, but I haven't seen it used much. I've used all sorts of different miniature flat things, which tend to go out of stock.

Small Transistors

This is what you would use to light up an LED that draws too much current for an ASIC pin, or to drive a small relay, or to level-shift a signal with an RTL inverter. These are SOT-23. Surface-mount versions of many of the old 2NXXXX parts are available, with identical pinouts and specs from many sources.

Microcontrollers

Anyone can argue PIC vs. MSP430 vs. AVR vs. etc. The PICs tend to come out cheaper than MSP430 or AVR, for similar features, and I find them easier to program in assembler. That's lucky, because the PICs are very close to hopeless to compile for. The PIC18s are somewhat better, but still just weird.

gcc ports exist for both MSP430 and AVR; I've used both, and never found a bug in either. The JTAG thing with the MSP430s always seemed a bit stupid to me, a quarter of the pins on the device, but now they have the Spy-by-Wire parts. It's nice that you can get five-volt-tolerant AVRs; that hasn't gone away yet.

I like the latest ARM parts. Digikey has the Sharp LH754XX, and I found those very nice to use. (These are not microcontrollers; you must wire them to flash and to RAM. They have all sorts of good peripherals on-chip, though, including many kinds of LCD controller, and the datasheets are extremely well-done.) I've used the Atmel AT91SAM7 ARM parts extensively, including their AT91SAM7X, with Ethernet; these put a lot of neat applications within reach. I have not been impressed with their datasheets or software support, though. ST has a comparable line of ARMs, which I have not investigated.

The LH75410 (and probably other parts in that `Bluestreak' line) has an interesting latchup issue. The chip runs at 3.3 V for the I/O, but 1.8 V for the core. There's an on-chip linear regulator, to make the core voltage from the I/O voltage. The regulator has an enable input; this permits the use of an external 1.8 V supply, if desired. If an external 1.8 V supply is used, then certain power supply sequencing restrictions apply; the core voltage should come up either before or only slightly after the I/O voltage.

My prototype boards worked, but production yields were poor; on many units, the processor did not appear to execute code, and drew a very large current, as if it were in CMOS latchup. After much wasted effort, I 'scoped the regulator's enable pin. This was pulled up to +3.3V through a 10k resistor, as specified in the datasheet. This should have enabled the regulator. In fact, that pin was somehow glitching low; the `input' was drawing current, and fighting the pull-up. As requested, the on-chip linear regulator shut down, the core voltage lagged the I/O voltage, and the processor latched up.

I `fixed' the problem by replacing the 10k with a zero-ohm, to fight the glitch. The most recent version of the datasheet also recommends this, but at the time that I designed the board, Sharp still wanted 10k. As an alternative fix, I noticed that a very fast-rising +3.3V rail (<100 μs) did not trigger the problem. That's not very practical, though, huge peak current to charge the decoupling caps.

My assumption is that Sharp found the problem, just as I did, and changed their datasheet accordingly. Considering the impact (~30% bad boards, including some that go bad in the field), it might have been nice of them to describe this somewhere prominent, instead of burying their defect in a footnote.

Integrated Circuits, General

I still like SOIC packages, if I have room for them. I usually don't, though. I avoid BGA and QFN packages when possible; prototype assembly and rework is straightforward with suitable equipment (hot air station), but very time-consuming. A lot of radio stuff isn't available any other way, so I often have no choice.

I like asymmetric packages; the SOT-23-5 is genius. I always use SOT-23 diodes, for the same reason. Surface-mount LEDs are particularly bad; is the dot the anode or the cathode? I thought there was a standard, but I have examples of both.

Train Wrecks

I had a terrible experience with the Freescale 802.15.4 radios (MC1319x; I tried both the MC13191 and MC13192). After a solid week of fiddling with temperature profiles, I finally got my prototypes to work, using a hot air station. I assembled about a hundred units using an IR reflow oven. About half worked. I ultimately blamed that on mechanical damage to the board during depanelisation. (Someone suggested that to me. I discovered that a working board could be converted to the non-working kind by flexing it slightly between my fingers. The solder joints are all still fine, though; no shorts, and I can still see the protection diodes on each pin with a multimeter, so no opens.)

I then gave up, attributing all of this to my incompetence with surface mount. I sent thirty units to two different commercial assemblers, already depanelised. At the end of this two boards worked. I sent the broken ones back for rework; of those, one worked. I gave up and switched to the Chipcon part, the CC2420. I ran a test batch of 24 units, and all of them worked. (All of these are QFNs, leadless packages.)

Freescale's tech support was actively useless. The assemblers that I dealt with gave me plausible suggestions that didn't fix my problem (moisture absorption, bad power supply sequencing, ESD, marginal timing, etc.). Their tech support just cut-and-pasted two paragraphs from the datasheet, and then marked my case `resolved.'

I also found a neat set of issues with the CYWUSB6934. This is another 2.4 GHz radio, this time nonstandard and from Cypress. I built some prototypes, and they worked, but only at long range: if I got the receiver too close to the transmitter, then it stopped working. I could make them work at short range by removing the antennas, but that of course broke them at long range.

The problem is with the AGC. Bit 7 in register 0x2e has changed with time. In rev E of the datasheet, it's `AGC Lock,' to lock the LNA to maximum gain. They say that `It is recommended to set this bit during initialization to save power.' In rev I, it's `Reg Power Control,' and `The application MCU must set this bit during initialization.' In an earlier rev of the datasheet that I can no longer find, they said it disabled the AGC, and that you should leave it clear.

So the AGC doesn't work, but they would rather not admit that. Instead, they want you to disable it. The problem with that is that at high input power levels (really high, like -50 dBm), something saturates in the receive chain, and the receiver stops working. This is presumably why they designed an AGC in the first place. This means that the `Maximum Received Signal' spec in Table 12-3 of the datasheet is unachievable; basically, it's a lie.

I spoke to their call center, which was completely unhelpful, but where they apologized from a script. I also spoke to an apps engineer in the US; he politely ignored my questions about Table 12-3. He did point me to their sample firmware, though. In the firmware that they provide, the transmitter does the AGC: it transmits a packet, and if it doesn't get an ACK, then it retries at lower power! This is the correct workaround for a part with a broken AGC, like this one. In my application I had a T/R switch between the radio and the antenna, so I just chose to flip it the `wrong' way, for a software-insertable ~20 dB pad.

If the datasheet just spec'd the lousy dynamic range, and made a note of it somewhere on the first page, then I would probably have still used the part; there's not a lot of competition. I'm not sure how it would have hurt them to publish an accurate datasheet.

Power Switches

I tend to use FETs. IRF makes most of the nicest parts, but there's some deals to be found elsewhere; something like the Toshiba 2SK2231 is well behind the curve, but very cheap.

If I use a PWM IC, then it will usually do the gate drive for me. Otherwise, I've done discrete gate drivers (complementary emitter followers, if you can afford to throw away 1.2 V), or Microchip has some useful `logic gates' with high-current output.

Once you get to reasonable current, BJTs seem worse, though not much so. A Zetex ZTX651, for example, gives similar performance to my 2SK2231 (Vcesat = 0.5 V max @ 2 A, Ib = 200 mA; vs. IRdson = 0.3*2 = 0.6 V, probably around 5 V gate drive and 2A), but costs 50% more, and you have to come up with that base drive somewhere.

BJTs seem to turn out better in high-speed, low-current applications. In that case it is not such a big deal to provide a few milliamps of base current, but it is difficult to build a cheap circuit to drive the FET's gate capacitance. It's possible to find (expensive) bipolar switches that still have decent Vcesat with a forced beta of a hundred.

A logic-level n-FET can be driven directly from an I/O pin through a large-value (kohms) resistor, if you don't have to switch it often. As long as the pin is always driven as an output, the transient currents will be clamped by the micro's output FETs, not by the protection diodes, which I am told is not nearly as bad (intuitive, I guess, not triggering the SCR).

Opamps

I've been impressed with Microchip's offerings; they seem to make good tradeoffs. I don't usually work with supplies beyond 5 V, or with negative supplies, so rail-to-rail (practically, complementary FET) outputs are essential. That's common, though.

Rail-to-rail inputs are less common, but Microchip seems to like them. Except for high-side current sense resistors, I'm not really sure what to do with them, but they don't usually hurt. The disadvantage is an increased input capacitance, and a nonlinear input offset voltage, which varies as a function of the common-mode input voltage.

The IC designer achieves a rail-to-rail input by using two input stages in parallel, one n-type, and one p-type. The p-type stage is used when the common-mode input voltage is near ground, and the n-type stage is used when Vcm is near the positive supply. In between, you get a mix. I'm not sure what happens when the differential input voltage is large (i.e., the opamp loses regulation), but it's probably bad.

If you look at the input offset voltage histograms for a rail-to-rail input opamp, then they will probably give you multiple figures, at different common-mode input voltages. The mean for all of them will be zero, because the process control people turn some knob to make that true. The standard deviation will be smallest when the opamp is biased in the middle of its supply; each input stage has its own independent offset voltage, and when you average two independent random variables, the standard deviation decreases.

as of Oct 2010, New York