Fullscreen HTML5Med Res (??MB)HTML4

Compact Frequency Divider Project by Nicholas Vinen This small board converts a standard 10MHz frequency reference (eg, from an oscilloscope) to 1MHz and 1Hz square wave signals. The latter can emulate the 1PPS output of a GPS receiver and has options for a 10% or 50% duty cycle. Build your own compact Frequency Divider 10MHz – 1MHz | 10MHz – 1Hz T his straightforward circuit accurately divides a 10MHz signal to 1Hz with extremely low jitter. It has various applications, such as testing clocks and other devices that are time-locked to GPS signals. It could also be used to drive several clocks from a single accurate time source or to derive a very accurate 1Hz signal from a low-cost 10MHz temperature-­c ompensated crystal oscillator (TXCO). The divider uses just four logic ICs, including the somewhat unusual 74HC4059, plus an ultra-high-speed comparator and a buffer for driving the outputs. It can be powered at 5V DC from a USB supply or 6.6-12V DC, drawing only about 10mA. It has supply reverse polarity protection and input overload protection and won’t generate an output unless it’s being actively fed a signal. A jumper selects between the 10% and 50% duty cycle options for the 1Hz output. The output jitter is extremely low as long as the input signal is relatively clean. Many pieces of test equipment will have a suitable 10MHz output, or you could use one of our GPS-­Disciplined Oscillators: • GPS-Disciplined Oscillator (May 2024) • GPS-synched Frequency Reference (October and November 2019) • GPS-based Frequency Reference (April & May 2009) Circuit details Its circuit is shown in Fig.1. We have tried to keep it simple and inexpensive without sacrificing performance. The 10MHz signal is fed into SMA connector CON1 and goes into the first stage, based on ultra-high-speed comparator IC6 (TLV3501). The TLV3501 is an interesting device as it runs from 2.7 to 5.5V, drawing just 3.2mA and yet has extremely low input bias currents at ±2pA (typical), a low input offset voltage of ±1mV (typical) and very high-speed operation with a maximum toggle frequency of 80MHz. That makes it suitable for many applications. Here, its job is to convert what might be a relatively low-level, sinusoidal input signal into a 5V peak-to-peak square wave. That means the circuit is not too sensitive about what drives it, as long as it is a 10MHz waveform of at least 10mV RMS or 35mV peakto-peak; it will most likely be a sine or square wave. The circuit is designed with 75W impedances in mind, although you could change that if necessary (eg, using 49.9W or 51W resistors instead of 75W). So the input is terminated with a 75W resistor, then coupled to comparator IC6 by a 1nF DC-­blocking capacitor and 220W series resistor. Dual schottky diode D1 protects IC6 from over-voltage or having a signal applied while the circuit is powered down by clamping the input signal to within 0.3V of the supply rails. Frequency Divider Features & Specifications » Divides the nominally 10MHz input frequency by 10 (to 1MHz) and 107 (to 1Hz) » 10% or 50% duty cycle option for 1Hz output (5V peak-to-peak unloaded) » Operating input signal level: 10mV to 3.2V RMS (28mV to 9V peak-to-peak) » Recommended input signal: 35mV to 2V RMS (100mV to 5.6V peak-to-peak) » Jitter: estimated at 0.1ns with a clean clock source (see Scope 1) » Propagation delay: approximately 100ns » High noise immunity with 23.5mV built-in hysteresis » Outputs are in phase with inputs » No output signals if the input is not driven » SMA connectors for input and outputs » Choice of 50Ω or 75Ω input/output impedances » Power supply: 5-12V DC <at> 10mA » Power connectors: USB Type-C, 2.1mm/2.5mm inner diameter barrel plug, (polarised) pin header » 3mm mounting holes: 4 (the board can be made smaller by cutting them off) Practical Electronics | March | 2025 The prototype board is very similar to the final version; it just lacks the power LED and used a Type-B mini USB socket instead of the now more standard Type-C. 39 Constructional Project Scope 1: the yellow waveform is the 1Hz output (reduced in amplitude due to a lower than normal supply voltage and 50W termination) while the blue waveform is the 10MHz reference signal from the oscilloscope. The grey areas around them show the previous 50 or so traces, indicating extremely low variation in timing between them (ie, low jitter). The output edge seems to come first because the propagation delay is just under the input signal period (100ns); it was triggered by the previous edge. The 220W resistor primarily exists to limit the current through these diodes, protecting them and the rest of the circuit from excessive ‘bus pumping’ of the 5V rail. As the signal is AC-coupled to IC6, it is DC-biased to half of the 5V supply using a pair of 10kW resistors and a 47kW bias resistor. The 100nF capacitor prevents supply ripple from coupling back into the signal, which could cause jitter. A 10MW resistor from output pin 6 of IC6 back to its non-inverting input, pin 3, provides around 23.5mV of hysteresis for noise rejection. This forms a voltage divider with the 47kW bias resistor for pin 3. With around 2.5V across the hysteresis resistor (regardless of whether IC6’s output is at 5V or 0V), 2.5μV flows through it and subsequently the 47kW resistor, causing an offset of around 11.75mV. That offset switches polarity as IC6’s output switches, meaning that any noise on the input signal would have to exceed 23.5mV to cause an unwanted edge at IC6’s output. It also means that there needs to be at least a 23.5mV peak-to-peak signal applied to IC6 before its output will start to toggle. Thus, it won’t oscillate without a signal at CON1. With the 75W termination resistor, plus the low-pass filter formed by the 220W resistor and IC6’s input capacitance (plus that of both diodes in 40 Scope 2: to check the frequency ratio was correct, we captured the unit’s output on the scope for two seconds and then measured the time between edges. Here three of the captured edges are overlaid, in yellow (-500ms), red (0ms) and green (+500ms). The yellow and green traces overlap, indicating they are exactly one second apart as per the scope’s timebase (and hence 10MHz reference oscillator). As we captured two full seconds of data, the time resolution is more coarse than in Scope 1. D1), the minimum signal the circuit will respond to is about 30mV peakto-peak at CON1. However, a higher level is recommended to ensure jitterfree operation. 30mV at CON1 implies a higher voltage at the signal source, probably closer to 60mV peak-to-peak. Frequency divider Now that we have a clean 10MHz square wave signal from IC6, it’s fed to the first divider, IC1. This is a 74HC4017 Johnson decade counter, a lower-­voltage, higher-speed version of the good old 4017 counter IC. These are inexpensive, run from 2-6V, operate at up to 77MHz with a 5V supply and provide ten 10% duty cycle outputs with different phase angles, plus a single 50% duty cycle output that’s phase-aligned with the input (and Q0 output). For IC1, we feed the 10MHz signal into the pin 14 clock input and get a nice 1MHz square wave from the Q5-Q9 output. The MR (master reset) line is tied low for constant operation, while the inverting clock input at pin 13 is also tied low as we are using the non-inverting clock input. The ten phase outputs, Q0-Q9, are not used in this case. The 1MHz output from pin 12 is fed to two places: firstly, to three of the six buffers in IC5 connected in parallel, then to the 1MHz SMA output (CON2) via a 75W impedance-matching resistor. The MC74VHCT50A is similar to a 74HC04 hex inverter IC except that it does not invert the signals but merely buffers them. That keeps the outputs in phase with the 10MHz input. Secondly, the pin 12 1MHz output of IC1 goes to another 74HC4017 counter, IC2, configured identically to IC1. It produces a 100kHz square wave at its pin 12 output, which is fed to the clock (CP) input, pin 1, of IC3. This is the ‘main event’, configured to divide its input frequency by a factor of 10,000. It is a larger IC than the others, with 24 pins rather than 16, and somewhat more expensive (but still pretty reasonable). It takes up less space than four more 74HC4017s and has a much lower propagation delay. It can be configured for thousands of different frequency division ratios in various ways based on the logic states of its KA-KC and J1-J16 pins. The accompanying panel explains how this particular configuration achieves the 10,000:1 division ratio. We could have added a microcontroller to this board, driving all those pins, and provided a few different ratios. However, we decided it was better to keep this simple and avoid programming any chips. Still, we will present a more complex programmable design that includes a microcontroller in an upcoming issue. Practical Electronics | March | 2025 Compact Frequency Divider Fig.1: the circuit uses three divideby-ten ICs (74HC4017) and one divide-by-10,000 IC (74HC4059) to reduce the 10MHz input at CON1 to 1Hz at CON3. High-speed comparator IC6 converts whatever waveform is fed in to a 5V peak-topeak square wave for driving IC1. IC3 has an output latch that we do not use, so the latch enable (LE) input, pin 2, is tied to ground. The 10Hz signal appears at pin 23 (Q). Note, though, that this pin will only be high for one input pulse, and with a 100kHz input, the output pulses are 10μs wide. That is why we divided the 10MHz signal by a Practical Electronics | March | 2025 factor of 100 first; otherwise, the output pulses would be a mere 100ns wide. To make this short pulse useful, we feed it to the final counter, IC4, another 74HC4017 configured much like the others. It performs the final division to get a 1Hz signal and converts the short pulses into a 50% duty cycle square wave at its pin 12 output. We feed that, plus the similar but shorter 10% duty cycle pulse from output Q0, to a three-way pin header. That allows you to select the desired duty cycle using a jumper shunt. The resulting signal is fed to another triple parallel buffer (IC5d-IC5f) and then the 41 Constructional Project final SMA output, CON3, via another 75W impedance-matching resistor. The 10% duty cycle output more closely simulates a GPS 1PPS output, while the 50% duty cycle signal is nice and symmetrical for driving something like a clock. Power supply There are three power supply inputs. The USB Type C connector (CON4) is the simplest as it feeds the USB 5V directly into the circuit. However, note that its ground connection goes via the internal switch in barrel socket CON5. This way, if you plug both in simultaneously, you won’t be feeding power into the device connected to the USB socket. Unlike USB Type-B sockets, the Type-C socket needs two 5.1kW pulldown resistors connected to signal the power source to deliver 5V. You can leave those resistors off the board if you aren’t fitting the Type-C socket. This particular socket only has the six pins needed for USB power delivery, without the data signals. By the way, we’re switching from Type-B to Type-C because it is now the universal standard, so expect to see more of this in future. After passing through CON5’s internal switch, the GND connection from CON4 also passes through Mosfet Q1 before reaching circuit ground. This provides reverse supply polarity protection, although that should not be necessary for the USB socket as the socket itself should guarantee the correct polarity. However, it is helpful if powering the circuit via barrel connector CON5 or header CON6. In those cases, as Q1’s gate is connected to the +5V rail and incoming DC supplies via two 10kW resistors, it will only conduct if the incoming supply polarity is positive. If it is negative, Q1’s gate will be pulled negative, Q1 will be off, and the whole circuit will be unpowered, floating at the positive DC supply voltage (that was erroneously connected to the negative input). There are two 10kW pull-up resistors for the gate so that Q1 will switch on regardless of whether the USB connector is used (feeding 5V directly) or one of the other inputs, which feed 5V Programming the CD74HC4059 counter This counter is quite complicated as it includes a prescaler plus a three or four digit ‘decimal’ main counter that varies in how you can use it. The prescaler can divide by between 1 and 10 in five different modes. However, which prescaler mode you choose affects what values you can have in the main counter’s top (thousands) digit. For example, if you have a divide-by-10 prescaler, the main counter only has three digits (up to 999). If you use one of the other prescaler values, the main counter has four digits, with more options as the prescaler division ratio becomes smaller. The lower three decimal digits of the main counter can always be preset with a value from 0 to 9. Depending on the mode, the overall maximum division ratio is either 9999 (eg, with the prescaler in divide-by-10 mode) or 15999 (with the prescaler dividing by a power of two). It is actually possible to divide by a much higher number than that because the ‘BCD’ or ‘binary coded decimal’ counter stages that it initially seems can only count up to 10 are actually full binary counters that can count up to 16. So, while programming it is trickier, it can be set to divide by up to 21,327. Luckily, our desired division ratio of 10,000 is relatively easy to set up. We could have used a prescaler value of 10, leaving a three-digit main counter. While dividing by 1000 with three digits seems impossible, we could have set the top ‘digit’ to 10 (because the actual limit is 15), which would have given the desired result. In the final design, we use a prescaler ratio of 8, leaving us with four digits for our main counter, although the top digit can only be 0 or 1. That’s fine because we set the main counter to divide by 1250, as 1250 × 8 = 10,000. The prescaler value of 8 is selected with KA low, KB low and KC high as per Table 1. We then program the top digit of the counter using J4, which we set high, to 1. The remaining three digits are set to 2, 5 & 0, as shown in Table 2. One ‘gotcha’ when setting up this counter is that, while the thousands digit for the counter is set using low-numbered inputs (J2-J4), the hundreds digit is set using the highest-numbered inputs (J13-J16). So the digits do not appear at the inputs in order, except in the mode when the prescaler can divide by up to 10. 42 low-dropout regulator REG1. Zener diode ZD1 prevents damage to Q1 as its gate is only rated to handle ±12V. This method has a much lower voltage loss than using a series diode (a few millivolts instead of 300mV+), allowing you to use a supply barely above 5V while still getting a regulated 5V at the output of REG1 to power the rest of the circuit. Construction While it uses mainly SMD parts, the board is relatively easy to assemble as they are all fairly large. Experienced constructors can gather the parts and solder them to the board as shown in overlay diagrams Figs.2 & 3. We suggest fitting all the SMD parts to one side of the board, followed by the other, then the through-hole parts. It’s best to start with the top, as more parts are on that side. The Frequency Divider is built on a double-sided PCB coded 04112231 that measures 64 × 37.5mm. We recommend soldering IC1-IC5 in numerical order, then ZD1, Q1, REG1, the USB socket (if fitting it), then the top-side capacitors and resistors. With the ICs, check very carefully that each one is the right way around before soldering them; most will have a pin 1 dot or bar. Use a magnifier to find them if necessary. As shown in Fig.2 and on the PCB, in each case, pin 1 faces towards the top of the board or to the left (for IC5). There is one 1μF capacitor on this side; the rest are 100nF. As mentioned earlier, you can leave off the 5.1kW resistors if you aren’t using the USB socket. Also note that unlike the TypeB USB sockets we’ve been using for a while, these Type-C sockets have no locating posts that slot into holes in the PCB, so you will have to be careful to align all its pins and tabs with the pads before soldering more than one. There are various ways to solder these parts: with solder paste and hot air, solder paste and a reflow oven, solder paste and a hot plate or regular solder and a regular iron (which is how we did it). If using a standard iron, we strongly recommend having a good quality flux paste on hand, plus some solder wick, as they make it much easier. There is no ‘right’ way to hand solder SMD ICs, but here is how we did it, starting with the ICs. We placed a little solder on one of their pads, then Practical Electronics | March | 2025 Compact Frequency Divider Figs.2 & 3: we recommend fitting all the SMDs on the top side first. Ensure all the ICs are orientated correctly and leave the SMA connectors, DC socket and headers until after you’ve populated the underside of the board. There are not as many components on the underside; just the comparator IC, passives and the dual diode. slid them into place while heating that solder (to keep it molten). Removing the iron, we checked that all the leads were centred on their pads. If not, we reheated that solder joint and gently nudged the IC towards the correct position, rechecking each time. Once the IC was positioned correctly, we soldered a couple more pins, then spread a thin layer of flux paste along both rows of pins, loaded the soldering iron tip with some solder and dragged it along the pins. Each one took up the right amount of solder, making quick work of all the joints. Only a few joints got too much solder, resulting in a bridge to an adjacent pin. We removed the bridges using a bit more flux paste and an application of clean solder wick. You could use a slightly different technique, where you clamp the device in the correct location using a clothes peg, haemostat clamp or similar, tack it down, then solder the remaining pins. That technique involves more set-up time but less trial-­and-error. Once the ICs are in place, you can solder the remaining three-lead and two-lead components with a similar technique. Just make sure you let one joint solidify (which can take a few seconds) before making the other, or you could end up pushing the parts out of position. With all the parts in place, clean the board with some flux cleaner (or pure alcohol if you don’t have a specific flux cleaner), let it dry and inspect all the solder joints to ensure you haven’t missed any imperfect/incomplete joints or bridges. Then flip the board over and solder the parts on the other side using a similar technique. There is just one chip (IC6) on the underside, plus one dual diode in a three-pin SOT-23 package and 10 passives (resistors & capacitors). Take care with the orientation of IC6; its pin 1 goes towards the nearest PCB edge. Some parts are close to IC6, so it’s best to solder IC6 first, then the components right next to it, followed by those further away. Again, when finished, clean off the flux residue and inspect your work. Finally, flip the board back over and solder the SMA connectors, the threepin header for LK1, plus whichever of CON5 and CON6 you will be using. If leaving CON5 off, you will need to solder the short wire link shown in red in Fig.2 and the PCB silkscreen, or the board won’t get power. Note that you could leave SMA connector CON2 off if you don’t need or want the 1MHz output. Testing The board should draw under 20mA when powered up. If you have a current-­limited bench supply, set it to 6V and at least 30mA and connect it to CON5 or CON6. If it goes into current limiting, switch it off and check for faults. If you don’t have a bench supply, use a regular DC supply fed through a DMM set to measure milliamps and switch off if the current shoots up when you power it up. Lacking such a supply, you just have to YOLO it: plug a suitable power supply in and check if LED1 lights. If it doesn’t, unplug the cable and try to figure out why. If it does, proceed with the following checks. Assuming the current draw is OK, check the voltage between the shell of one of the SMA connectors (ground) and the large tab of REG1. It should be close to 5V. If it is below 4.75V or above 5.25V, check the soldering on REG1 and its adjacent bypass/filter capacitors. If it isn’t drawing any current and the LED is off, that probably means that Q1 is not conducting. You can verify that by measuring the voltage between Table 1 – 74HC4059 modes (● must be set up with Master Preset mode first) KA KB KC Prescaler ratio Preset inputs Counter thousands digit Preset inputs Maximum count 1 1 1 2:1 to 1:1 J1 0-7 J2-J4 15,999 (17,331 extended) 0 1 1 4:1 to 1:1 J1, J2 0-3 J3, J4 15,999 (18,663 extended) 1 0 1 5:1 to 1:1 ● J1-J3 0-1 J4 9,999 (13,329 extended) 0 0 1 8:1 to 1:1 J1-J3 0-1 J4 15,999 (21,327 extended) 1 1 0 10:1 to 1:1 J1-J4 0 - 9,999 (16,659 extended) Table 2 – our 74HC4059 configuration KA KB KC Prescaler preset (J1-J3) Thousands (J4) Hundreds (J13-J15) Tens (J9-J12) Units (J5-J8) 0 0 1 000 (0) 1 (1) 0101 (5) 0000 (0) Practical Electronics | March | 2025 0010 (2) 43 Constructional Project your supply negative and the shells of the SMA connectors. There should be very little difference. If you measure the full supply voltage, check that you’ve applied power with the correct polarity. If you have, there is a fault around Q1/ZD1. Finally, assuming the current draw is OK and the 5V rail is close to 5V, feed a signal with a known frequency into CON1 and check for 1/10th that frequency at CON2 (if you didn’t fit CON2, you can probe its centre pin). If that checks out, apply 10MHz to CON1 and look for a 1Hz output at CON3. If it’s missing, make sure JP1 is inserted in one of the two possible positions. Remember that, depending on your test instrument, it could take several seconds to register a reading of such a low frequency. If the board isn’t behaving, common problems to look for are solder bridges, pins where the solder hasn’t adhered to the PCB pad below, or incorrectly orientated ICs (we did warn you!). Usage There isn’t much to it: connect your reference signal source to CON1 and feed the output at CON3 to your GPS clock(s) or other devices needing 1Hz pulses. Move JP1 if necessary to get the desired duty cycle, although almost any device expecting a 1PPS signal should work in either position. We suggest housing the board in a small diecast aluminium box with the case connected to circuit ground to minimise EMI pickup. However, we tested it as a ‘bare board’ and it performed well in our lab. The SMA connectors are arranged along one edge, so you can mount the board such that they project through holes in the case, then add a chassis-mounting DC socket wired to CON6. The four corner mounting holes will provide a convenient way to attach the board to the inside of such a box. If you need to make the board as small as possible, the tabs those holes are on can be cut off with a hacksaw or similar (but don’t breathe the resulting dust Parts List – 10MHz Frequency Divider 1 double-sided PCB coded 04112231, 64 × 37.5mm 3 right-angle or vertical through-hole SMA connectors (CON1-CON3) 1 SMD USB Type-C power-only socket with six pins (CON4) ● 1 PCB-mount DC barrel socket (CON5) ● 1 2-way polarised header, 2.54mm pitch (CON6) ● 1 3-pin header, 2.54mm pitch (JP1) 1 jumper shunt (JP1) ● omit any of these power input connectors that are not needed Semiconductors 3 (CD)74HC4017(M96) high-speed CMOS Johnson decade counters, narrow body SOIC-16 (IC1, IC2, IC4) 1 (CD)74HC4059 high-speed CMOS programmable divide-by-N counter, wide body SOIC-24 (IC3) 1 MC74VHCT50A hex CMOS non-inverting buffer, SOIC-14 (IC5) 1 TLV3501AID rail-to-rail high-speed comparator, SOIC-8 (IC6) 1 AMS1117-5.0 or compatible 5V 1A low-dropout regulator, SOT-223 (REG1) 1 AO3400 30V 5.8A N-channel logic-level Mosfet or equivalent, SOT-23 (Q1) 1 SMD LED, SMA/M3216/1206 size, any colour (LED1) 1 BZX84C5V6 5.6V 1% tolerance zener diode, SOT-23 (ZD1) 1 BAT54S dual series schottky diode, SOT-23 (D1) Capacitors (all SMD M3216/1206 size 50V X7R) 8 100nF 1 1nF 1 1μF Resistors (all SMD M3216/1206 size 1%) 1 10MW 1 47kW 4 10kW 2 5.1kW (only needed if USB socket is fitted) 1 1kW 1 220W 3 49.9W, 51W or 75W (to suit desired input/output impedance) 10MHz Frequency Divider kit from Silicon Chip (Cat SC6881): ~$48 (£24) + P&P 44 and cut them outdoors or in a wellventilated area). Most oscilloscopes, spectrum analysers, high-end frequency counters etc will have a pretty accurate 10MHz output; it’s usually specified as something like ±1ppm. That isn’t as good as a GPS-disciplined oscillator but it’s still very precise. You will likely need a BNC-to-SMA cable to make this connection. You may need a second similar cable for the 1Hz output, depending on where it’s going. Lacking that, some newer DSOs have a waveform generator output that can generate a 10MHz sinewave or square wave (either is suitable). They tend to have quite a bit less stability and more jitter than a 10MHz reference output. However, an actual GPS 1PPS signal has jitter, so if you are using this board to emulate such a signal, you generally needn’t worry too much about it. You can also get connectors that break a BNC connection out to screw terminals if you’re going to feed the 1PPS signal to pin headers or similar. If you want to feed the 1Hz output of this board to multiple clocks or other devices, given its low frequency and the fact that most 1PPS inputs will have a high impedance, you will probably just need to ‘fan it out’. You could even omit CON3 and solder wires directly to its pads. We mainly provided the SMA connector for convenience in hooking it up to prebuilt test equipment. If you need to split the 10MHz output of your test equipment to go to multiple locations, consider building our Frequency Reference Signal Distributor (April 2021). However, note that the design won’t work on the 1Hz output without modification as it is AC-­coupled at the input PE and outputs. We used rightangle SMA connectors. Practical Electronics | March | 2025 AI-enabled holograms Techno Talk Only a few years ago, most of us would have thought the technology behind life-sized AI-enabled human holograms was far away, in the distant future. Well, tomorrow must have come early, since I just saw people holding a conversation with such a beast! I feel like I’m straddling multiple technological epochs and I’m not as limber as I used to be. For example, I was just reading a very interesting 6-part series on the Rise and Fall of Heathkit (https://pemag.au/link/ac3c) written by my friend Steve Leibson. The 1960s and 1970s were golden years for Heathkits. These ranged from entry-level projects like building a crystal radio or an oscillator to practice Morse code, all the way up to advanced kits like ham radio transmitter/receivers and colour televisions. As Steve says, “Many engineers started their budding careers by building one or more kits made by the Heath Company”. I remember those far-off times as if they were yesterday. Meanwhile, I currently maintain a tenuous toehold in the present, in which artificial intelligence (AI) and machine learning (ML) applications are rampaging across the technological landscape. Things are currently moving incredibly fast in AI/ML space, where no-one can hear you scream. For example… Deploying AI at the edge It’s common to hear the term ‘edge’ used in the context of embedded systems and the internet of things (IoT). This means different things to different people, but I’m talking about what I think of as the ‘extreme edge’, that is, the point where the internet ‘rubber’ meets the real-world ‘road’. We are currently sitting on the horns of a dilemma, and it’s jolly uncomfortable, let me tell you. The crux of the problem is that almost everyone wants to add AI/ML to their embedded/edge applications and devices, but very few people have the necessary expertise to add AI/ML to their embedded/edge applications and devices. Just saying “crux” causes (what I laughingly refer to as) my mind to reflect on Frank Zappa’s playful and enigmatic line: “The crux of the biscuit is the apostrophe”, from his 1974 album Apostrophe (‘). This captures Zappa’s genius for making Practical Electronics | March | 2025 the mundane seem profound and the profound seem mundane. But we digress… Recently, I had quite the fascinating conversation with the folks at DeGirum (degirum.ai). These chaps and chapesses have a cunning solution to this problem. First, they have a “Model Zoo” that contains 1,000+ pre-trained AI/ML models ranging from things like people detection, face detection, age estimation, gender classification, and emotion classification (happy, sad, angry…) to vehicle detection, number (license) plate detection, pothole detection, fire and smoke detection, intruder detection and weapon detection. These models, which can be used in isolation or in combination, address the needs of a vast range of markets and applications. That includes ‘smart’ cities, offices and homes; construction, infrastructure and industrial automation; retail analytics and digital signage… the sky’s the limit. The folks at DeGirum also have a hardware-agnostic PySDK (Python software development kit). The clever bit is that they have a collection of supported hardware platforms ‘in the cloud’. Developers can combine their application software with DeGirum’s pre-trained models and run everything remotely on different platforms to determine the optimal cost/performance tradeoffs. Best of all, this is free! People pay based only on the number of PySDK runtime instantiations that are eventually deployed in real products in the field. Closer than we think Did you ever see the 2002 movie interpretation of H. G. Wells’ The Time Machine? There’s a scene set in a futuristic library where our hero converses with an AI-enabled hologram. You can find this moment on YouTube (https://youtu.be/CQbkhYg2DzM). When I first saw this movie (the day it came out), this level of technology seemed to be something that was a long, long way in the future. Max the Magnificent Well, I was just chatting with Edward Ginis, who is the chief technology officer (CTO) at a company called Proto (protohologram.com). They essentially have this technology in the here and now! Here’s how it works in a crunchy nutshell. First, they video someone talking for a few minutes. Let’s say that someone is me. Based on this, their AI can extract my ‘vocal signature’ for use in a text-to-speech role. It can also determine how my lips and the muscles on my face interact with the various spoken sounds. This includes the way I smile, frown, blink etc. It’s also watching the way I move my hands and arms; my entire non-vocal communication or body language. This is where things get extremely clever. Suppose you were lucky enough to be having a live conversation with me. While you were talking, although it wasn’t my turn to speak, I would still be engaged in the conversation with my body language: smiling, frowning, nodding in agreement… that sort of thing. Also, my mind would be constantly evaluating my potential responses based on what you were saying. Now suppose that you were talking to Proto’s AI hologram version of me. While you were talking, the hologram would be performing its non-verbal communication activities. Also, it would be deciding how to respond. This all happens in real-time (not like today’s Alexa-like assistants that wait for you to finish talking, then upload what you’ve said, and then reply after a few seconds’ delay). This means the AI hologram will respond immediately—sometimes before you’ve even finished talking— just like in a real conversation. Also, all visual aspects of the AI hologram—lips, facial muscles, hand and body gestures—are synchronised with whatever it’s saying. And just what will it be saying? I’m glad you asked. I’ll tell you about that in my next column. PE 45