Why use 2N3904 vs 2N2222?

[schmitt trigger]

There is a reason they are called “jellybean” transistors.

[rfclown]

I would say that the only reason is price. I just looked on Digikey, and was surprised to find that doesn't seem to be the case at this moment.

Comparing ON semi:
mmbt3904 Qty 100 $0.0496, Qty 1000 $0.02703
mmbt2222 Qty 100 $0.0536, Qty 1000 $0.02926

with that, I'd only use 2222 if all other variables were the same. I did a design last year at work in which I only needed two jelly bean NPNs (low current voltage doubler). I saw that we already had a mmbt3904 in our system, so I used that (it's work to put a new part in the system). I breadboarded the circuit with leaded parts to prove the design, and put SMT on the schematic. That's something nice about 3904 and 2222; you can get them in leaded to breadboard, and SMT to produce. It used to be that the price was different enough that if you didn't need the 2222, you'd use the 3904. It seems at the momemt that isn't the case.

In my case, the point about being already in the system pretty much trumps anything else. I also wanted to use two Schottky diodes in my circuit, so I found suitable diodes already in our system. If my object was really low cost, and I didn't care about the extra voltage drop, I could have used another two mmbt3904s as diodes and had less parts on the BOM. For this job, cost was absolutely no consideration whatsoever. Getting it done fast, and having a design that worked first pass was paramount.

[T3sl4co1l]

It's quite possible to use generic transistors in... most anything below VHF.  The main differentiators are overall ratings, hFE, and testing.

Ratings:
- Ic and capacitance are ~proportional.
- fT tends to decrease with higher current ratings (except for special types).
- Capacitance decreases some with increasing Vceo.

Testing separates parts by hFE, Vceo and noise:
- Within a process, hFE varies over a pretty wide range, so there's test selection within one process already.  (Different processes can achieve different ballparks of hFE.)
- Similarly, Vceo varies, so can yield a range of parts after testing (e.g. TIP31A/B/C).  This is more reliable these days (i.e. since the 70s, I suppose) so you don't see consecutive parts in a series of Vceo's so much anymore.
- Noise depends on process and purity, and can vary from part to part, so good parts can be tested out as well.  Not that the noise is necessarily particularly good on those parts, anyway; perhaps flicker noise is tested out, but r_bb' might still be objectionably high, limiting the wideband noise factor.

What you definitely won't get, is bandwidth or noise performance in a given power budget.  2N2222 or 2N4401 is a much larger junction than 2N3904, so you need that much more bias current, to achieve the same bandwidth.

And neither one has terrific fT.  Enough to scratch the bottom edge of VHF, but certainly not even enough to oscillate in UHF, let alone do useful work up there.  (Aside from peculiar operating modes, like avalanche breakdown -- which can produce sub-nanosecond edges, but isn't proportional in any meaningful way, and definitely doesn't have the bandwidth, as the pulse repetition frequency is sub-MHz).

Not that junction size is, by itself, a contraindication of noise performance -- indeed, large junctions act to spread out random fluctuations.  Low noise op-amps use extremely large input transistors (often taking up 25% or more of die area!).  Some particularly low noise transistors (mind, in terms of e_n) are rated for several amperes (a few ZTXxxx parts, a few power transistors -- but again, they must be tested and selected for low noise), and this can be had even at relatively low Ic (say, 10s mA, compared with the 1A+ rating).

All a matter of what you're doing.

Tim

[rfclown]

A few months ago I ran across a circuit which uses the negative resistance of a jellybean NPN to form a one transistor RC oscillator (without resorting to a unijunction transistor).

http://cappels.org/dproj/simplest_LED_flasher/Simplest_LED_Flasher_Circuit.html

I immediatedly breadboarded the circuit with parts that were laying in a pile on my bench... including a 2n3904. It worked. Then I tried it with a 2N2222. It worked much better.

Then I wondered about another trick circuit, a Bob Pease tease of using reverse breakdown in an NPN to generate a negative voltage. I had always shown the trick with a 3904. It worked better with a 2222. After that I ordered 30 PN2222s on eBay thinking that my go-to jellybean NPN should maybe be a 2222 instead of a 3904 (which it had been for over 40 years).

[mawyatt]

I was talking with the late Jim Williams at the IEEE ISSCC conference the year he passed away and he discussed this very type reverse bias avalanche breakdown effect. He had cleverly created a cheap low TC reference diode utilizing a revised reversed biased NPN transistor EB junction and passing the current thru the BC junction. So the transistor had a +V applied to the emitter, base open, and collector grounded. I can't remember the transistor used but recall it had a EB breakdown of 6~7V with a +TC of ~ +2mv/C, and the BC junction forward biased with a ~ -2mv/C, thus the emitter voltage was the EB voltage (6~7V) plus the BC junction (~0.6V) and had ~ 0 TC. He indicated that LT built a bunch of the PCBs for a client utilizing this as a cheap reference diode only to find out that the voltage was 6.6~7.6V average but had a small sawtooth ripple waveform caused by this avalanche effect on the EB junction!!

Even the best of the best get bit once in awhile by Murphy!!

Edit: Just tried what Jim Willams had done but with the 2N3904. Used a 1K series current limiting resistor in series with emitter. Slowly increased power supply voltage until ~10 volts  was reached, then the avalanche effect was observed with the reversed biased emitter voltage dropping to ~8 volts (@~2ma). Voltage remained @~8.2 volts with supply increased to ~15 volts (current ~7ma). When PS was dropped back to below ~9V (<1ma) then emitter returned to high impedance state @ ~0ma.   

Since this discussion is about 3 terminal solid state devices, there was a post in the IEEE Electronic Letters 30 or more years ago about using a standard 3 terminal solid state device that has an AC small signal voltage gain greater than unity over a moderate frequency range, but has no external components, no inductors, capacitors, resistors nor transmission lines, and no bias voltage nor currents. I used to give this my grad students in RFIC design, even tho it wasn't a typical RF circuit, to try and get them to think "outside the box". Try to think about this before running to Google (if it's even there I don't know). As a hint, there's a version you can simulate and build with ordinary Rs and Cs to mimic the 3 terminal solid state device.

Best,

[rfclown]

...
Since this discussion is about 3 terminal solid state devices, there was a post in the IEEE Electronic Letters 30 or more years ago about using a standard 3 terminal solid state device that has an AC small signal voltage gain greater than unity over a moderate frequency range, but has no external components, no inductors, capacitors, resistors nor transmission lines, and no bias voltage nor currents. I used to give this my grad students in RFIC design, even tho it wasn't a typical RF circuit, to try and get them to think "outside the box". Try to think about this before running to Google (if it's even there I don't know). As a hint, there's a version you can simulate and build with ordinary Rs and Cs to mimic the 3 terminal solid state device.

Best,

Haven't been able to figure this one out. Only thing I've thought of (and I know it's not the answer): If you took a 3 terminal linear regulator, like a LM317, and fed a signal into the adj pin, you'd have LOTS of AC gain (and DC). But it would have to have a DC offset exactly equal to the reference voltage or it would rail. And that DC offset, I would have to consider a bias voltage (and you could never do it except in a perfect simulation).

[mawyatt]

...
Since this discussion is about 3 terminal solid state devices, there was a post in the IEEE Electronic Letters 30 or more years ago about using a standard 3 terminal solid state device that has an AC small signal voltage gain greater than unity over a moderate frequency range, but has no external components, no inductors, capacitors, resistors nor transmission lines, and no bias voltage nor currents. I used to give this my grad students in RFIC design, even tho it wasn't a typical RF circuit, to try and get them to think "outside the box". Try to think about this before running to Google (if it's even there I don't know). As a hint, there's a version you can simulate and build with ordinary Rs and Cs to mimic the 3 terminal solid state device.

Best,

Haven't been able to figure this one out. Only thing I've thought of (and I know it's not the answer): If you took a 3 terminal linear regulator, like a LM317, and fed a signal into the adj pin, you'd have LOTS of AC gain (and DC). But it would have to have a DC offset exactly equal to the reference voltage or it would rail. And that DC offset, I would have to consider a bias voltage (and you could never do it except in a perfect simulation).

If you have a depletion mode semiconductor device such as a JFET or GaN FET and drive the gate terminal with a small voltage signal and ground the source terminal, then "look" at the output voltage signal at the drain terminal with a high impedance instrument. Sweep the input frequency and the output signal voltage will exceed the input gate terminal voltage at a moderate band of frequencies. This is NOT because of inductive resonate effects, nor transmission line effects, but because of the distributed capacitive coupling between the gate and channel which with a depletion mode device the channel is highly conductive with zero DC bias, voltage and current on all terminals!! A discrete version of this connect a number of resistors in series, with one lead of the string to ground and the other end of the string is the output, this emulates the channel. Now with an equal number of capacitors connect one lead of each capacitor to the individual resistor connections, and all the other cap leads together, this emulates the gate.

Here's a quick Spice sim thrown together of the 2N5434 JFET and an attempted RC network approximation (still needs some work). As you can see the gain is about ~+0.4dB @ ~2GHz. If you make R9 zero in the RC network you can achieve a higher gain, this is to help model the JFET gate resistance.

There's another version of the RC network that you can't do with the depletion mode device, you interchange the Rs and Cs. About 40 years ago we build these RC networks with probably 1K resistors and 0.1uF caps for a lower frequency where gain > unity, they work

Best,

[exe]

This is NOT because of inductive resonate effects, nor transmission line effects, but because of the distributed capacitive coupling between the gate and channel which with a depletion mode device the channel is highly conductive with zero DC bias, voltage and current on all terminals!!

Does this means that we simply inject charge into source-drain? If so, why does this have gain?

[mawyatt]

This is NOT because of inductive resonate effects, nor transmission line effects, but because of the distributed capacitive coupling between the gate and channel which with a depletion mode device the channel is highly conductive with zero DC bias, voltage and current on all terminals!!

Does this means that we simply inject charge into source-drain? If so, why does this have gain?

It has a voltage gain but not power gain which would violate conservation of energy!

The reason for voltage gain is the distributed charge from the gate along the channel creates a channel potential that slightly adds vectorially along the channel over a small range of frequency. With the discrete RC version you can see this by probing each RC node starting at the bottom near ground and look at the voltage magnitude and phase, as you work towards the output these individual voltages across each resistor add one on top the other in a fashion to produce the gain at the output node. You can swap the Rs and Cs and get a similar result (gain >1 but not at same frequency range), however this becomes a low pass function rather than a high pass function.

These circuits are easy to build on a plug in board with larger caps & resistors to keep the frequency lower, just account for the impedance of the output measuring device.

Many fun circuits around to play with. Another interesting circuit that had a practical use back in the 60s was using a transistor upside down, basically interchanging the collector and emitter. Running upside down the device has a reverse beta that's lower than the normal forward beta but still useful. These were often used as analog switches since the saturation voltage was very low, much lower than froward saturation.

Best, 

[T3sl4co1l]

Many fun circuits around to play with. Another interesting circuit that had a practical use back in the 60s was using a transistor upside down, basically interchanging the collector and emitter. Running upside down the device has a reverse beta that's lower than the normal forward beta but still useful. These were often used as analog switches since the saturation voltage was very low, much lower than froward saturation.

Incidentally, this is still true of modern low-Vce(sat) types -- I measured a PBSS303NX I think at 50-75 inverted hFE.  Vebo nothing to write home about of course, but it makes sense; that's why BR is a parameter in the Gummel-Poon model.

Oddly, the models for these parts, tend not to have a representative BR.  Wouldn't be surprised if they're simply ignoring that range of operation, when fitting the models.

There are still one or two "muting" transistors out there, I think?  These still aren't symmetrical types, regular planar diffused; but they manage to get useful Vebo (>12V?) and high inverted hFE.  They perform better than NMOS switches, I think? -- in terms of R(on) vs. C(off) at a given V(off,max), that is.  The downside is wiring up the bias currents to control them.

Tim

[rfclown]

...
Here's a quick Spice sim thrown together of the 2N5434 JFET and an attempted RC network approximation (still needs some work). As you can see the gain is about ~+0.4dB @ ~2GHz. If you make R9 zero in the RC network you can achieve a higher gain, this is to help model the JFET gate resistance.
...

I was wondering if I could measure this. LTspice sim of the 2N5434 said that even with 50 ohms on the drain you'd still have 0.12 dB gain at 2.8 GHz. Cool. Unfortunately, any series resistance at the input kills it. My ESG doesn't have the 0 ohm source resistance that I'd need.

[mawyatt]

...
Here's a quick Spice sim thrown together of the 2N5434 JFET and an attempted RC network approximation (still needs some work). As you can see the gain is about ~+0.4dB @ ~2GHz. If you make R9 zero in the RC network you can achieve a higher gain, this is to help model the JFET gate resistance.
...

I was wondering if I could measure this. LTspice sim of the 2N5434 said that even with 50 ohms on the drain you'd still have 0.12 dB gain at 2.8 GHz. Cool. Unfortunately, any series resistance at the input kills it. My ESG doesn't have the 0 ohm source resistance that I'd need.

Might be difficult to measure at those frequencies accurately without some expensive equipment. However if you build the RC version and use 1K resistors and 0.1uF caps you get this result, which is easy to measure with a good DMM. Note the "Voltage Gain" is ~1.1dB @ ~200Hz.

Best,

[mawyatt]

Many fun circuits around to play with. Another interesting circuit that had a practical use back in the 60s was using a transistor upside down, basically interchanging the collector and emitter. Running upside down the device has a reverse beta that's lower than the normal forward beta but still useful. These were often used as analog switches since the saturation voltage was very low, much lower than froward saturation.

Incidentally, this is still true of modern low-Vce(sat) types -- I measured a PBSS303NX I think at 50-75 inverted hFE.  Vebo nothing to write home about of course, but it makes sense; that's why BR is a parameter in the Gummel-Poon model.

Oddly, the models for these parts, tend not to have a representative BR.  Wouldn't be surprised if they're simply ignoring that range of operation, when fitting the models.

There are still one or two "muting" transistors out there, I think?  These still aren't symmetrical types, regular planar diffused; but they manage to get useful Vebo (>12V?) and high inverted hFE.  They perform better than NMOS switches, I think? -- in terms of R(on) vs. C(off) at a given V(off,max), that is.  The downside is wiring up the bias currents to control them.

Tim

The Gummel-Poon model is generally inadequate at accurately modeling higher current effects, even in the forward direction. Even VBIC doesn't do a very good job, we had to resort to using HICUM and MEXTRAM models in critical circuits and used VBIC in sections that weren't as critical like biasing and such. HICUM and MEXTRAM are very compute intensive, and often caused crashes, or ridiculously long simulations. Fortunately Cadence allowed the use of different models (box selectable from the schematic) within a given circuit simulation and IBM had created various type models for the devices/processes we were using, and we would only use the higher fidelity models where needed.

Speaking of models, back in 2008~9 when we staring working with the PolyPhase Mixer (PPM) we discovered that the Cal Berkley BSIM4 MOS model had a numerical error. The 1st derivative of channel transconductance has a discontinuity at a drain current of zero, this was due to the computation of the gate to source and gate to drain voltage when the drain current went thru zero like when the MOS FET is used as a switch with bidirectional current flow. This error is caused by using the absolute value of the gate to source/drain voltage to satisfy the device computations and creates a higher than actual channel modulation non-linearity. A solid-state physics based MOS model called PSP (Penn State/Phillips) didn't have the issue since it's computations were based upon channel surface potential rather than G-S-D terminal voltages, and we used this PSP model where needed like with the bipolars. Since Berkley didn't want BSIM4 or later versions to be a physics based due to the extreme complexity, they needed to find a "fix" that didn't create additional complications. Without having to resort to some arbitrary type solution, or some trickery, they came up with a brilliant solution by simply referencing the gate, drain and source voltages, not to each other, but to the channel substrate. This allows the interchange of the drain and source as in under switch bi-directional current flow use without any drawbacks....clever indeed

We looked at all sorts of solid state switches back then for PPM use, and found a simple large W/L NMOS of the smallest feature size (standoff voltage wasn't an issue) creates the most effective solid state switch from a gate drive vs. Ron vs. capacitance capability, and produced the lower odd order switch non-linearity with triple well NMOS, if available. Later when FinFETs arrived, they proved even better due to the higher effective gate to channel coupling. We never looked at bipolars since current biasing the switch was not an option. The PPM required a fast linear bi-directional unbiased channel switch to produce the "breaking the laws of mixer physics" performance.

Best

[rsjsouza]

Interesting; when I started in electronics, the 2N2222 was expensive and we considered it a specialty great for higher current and switching. Our jellybeans at the time were the BC107/108/109, coming just after the BC327/337 pairs. Only a while later I settled on the common BC547/BC548/BC549 and their corresponding PNPs. If you needed more gain, just use one with a suffix A, B or C.

[mawyatt]

The original 2N2222 and complement 2N2907 were in hermetic T0-18 cases and yes relatively expensive. When the transistor die were placed in the cheaper plastic packages like the TO-92 they were much cheaper.

Speaking of how cheap a transistor today is, consider the cost of a transistor in an IC! For example, take a typical USB FLASH memory. You can buy these for ~$10 for 64GB. Each MOS memory transistor may support 2 or 3 bits of data, so 3 or 4 transistors per byte. So at least ~ 3*64e9 MOS transistors at ~$10 cost, that's pretty darn cheap

50 years ago would anyone have thought a transistor would be cheaper than a resistor, capacitor, wire, washer, nut, or bolt, and possibly even cheaper than a kernel of corn or grain of rice

Best,

[T3sl4co1l]

70 years ago, perhaps not.  The planar IC had just been introduced, many might not've appreciated its potential (or even been skeptical of its success).

50 years ago was only 1970: ICs were already dominant and the cost per transistor (integrated) already well on its asymptotic course for the then-forseeable future!

Tim

[mawyatt]

Moore's Law was in it's early infancy 50 years ago, I know was there then and remember it well. Few, if any, probably Moore included, would have thought about nanometer level devices then. 30 years ago many knowledgable semiconductor folks couldn't foresee going much below a micron since photolithography imposed a theoretical limit to processing and e-beam was impractical (take's too long) for mass production. Even 20 years ago, after diffraction limits was being pushed aside, 5nm transistors weren't envisioned by many, myself included

What we have witnessed with silicon semiconductor technology in the past couple decades is absolutely remarkable, if Einstein were with us today, he would likely say "You Can't Do That!!" 

And equally remarkable IMO is that we have these impossibly small devices in mass production inside our very phones and computers we use everyday, not some future products in 2040!! A modern chip design in a leading edge fab cost over $500M, the fab construction cost over $20B, and yet we have these chips at modest costs in our present-day products. I mean, chips for a few tens of bucks, from a non-recurring engineering cost of ~$500M, produced from a $20B fab.....the economic side of this is also staggering!! Of course the answer is massive production scales

So the message is enjoy your transistors and say thanks to the semiconductor folks that make it happen

Best,

[CJay]

There are two versions of the PN2222 as well with different pinouts 

I had a bad day when I encountered this.  Old ones had C & E swapped compared to new ones.  They still "work" (BJTs are somewhat symmetric) but very poorly and often go pop. 

I now ignore the datasheets for all TO-92 BJT parts and use a transistor tester to tell me the pinout instead.   It also doesn't help when the odd datasheet doesn't tell you if it's a top or bottom view of the pinout

Pain in the fundament that, I've been bitten by it when a supplier had mixed batches up.

It doesn't just apply to the PN2222 either, mine were BC5something or other so I now read the datasheets then stick it on the transistor tester.

(which reminds me, I need to order up a couple of ATMegas after a depressingly unspectacular incident with a 42,000uF cap).

[SilverSolder]

There are two versions of the PN2222 as well with different pinouts 

I had a bad day when I encountered this.  Old ones had C & E swapped compared to new ones.  They still "work" (BJTs are somewhat symmetric) but very poorly and often go pop. 

I now ignore the datasheets for all TO-92 BJT parts and use a transistor tester to tell me the pinout instead.   It also doesn't help when the odd datasheet doesn't tell you if it's a top or bottom view of the pinout

The old maxim, "trust, but verify", very much applies to data sheets!