Wanted a track light fitting for the living room and more aluminium welding practice 🙂
20x20mm square tube (ashamed to admit I bought it from B&Q) was cut into the shape of a lightning bolt.
Oh dear, the mitreing could have been better…
This is the nicest weld, there were much worse 😀 I got off to a bad start by burning several holes in it while trying to tack it together, and had to weld up the holes. I didn’t realise the tubing was only 1.6mm thick and started off with too large a tungsten and too much current. Another reminder to always test the welder settings on a scrap of the material you’re going to use… 🙂
After fully welding (OK I didn’t do the inside corners 🙂 )
As you can imagine from the number of acronyms in the title, this took a huge amount of fiddling and was very confusing, so documenting it here for my own benefit and others’. 🙂
I wanted to run Ubuntu Server ARM64 on a Raspberry Pi CM4 dev board using a NVMe SSD for at least the root filesystem (and ideally booting from it)
Step 1: Update the CM4’s bootloader firmware
The CM4 probably has an older bootloader firmware installed from the factory that doesn’t support NVMe. This step is best done from a Linux PC: I used an old Acer laptop running Zorin 16. The below commands are to be entered in a terminal window:
First you need the latest version of the Raspberry Pi USB boot tool. This has to be built from source, and you may need to install some dependencies first (-1000 geek points if you didn’t already have git and build-essential 🙂 )
Now it’s time to build the USB boot tool from source…
git clone --depth=1 https://github.com/raspberrypi/usbboot
Now you need to edit the boot order settings and get the latest EEPROM image. This part is adapted from this tutorial which doesn’t work verbatim any more (the “nvme” folder doesn’t exist in the usbboot repository, but substituting the “recovery” folder works)
sed -i -e '/^BOOT_ORDER=/ s/=.*$/=0xf25416/' boot.conf
NVMe on Raspberry Pi is still kind of bleeding edge, so it’s best to use the newest versions of everything. pieeprom-2021-07-06.bin was still the latest image at the time of writing:
Note: At this point the update script crashed complaining that it couldn’t find Python. I guess Zorin 16 removed Python 2.x and forgot to make Python 3 the default? The python-is-python3 package fixed that for me. Hopefully your Linux distro has a functioning “python” already, in which case don’t execute this command, it’ll probably break something!
sudo apt install python-is-python3
Finally it’s time to burn your changes to the CM4… The “disable eMMC Boot” jumper must be fitted on J2 for this to work (Unless you have a CM4 Lite I guess…)
sudo ./rpiboot -d recovery
Now connect the CM4 Dev board J11 “USB Slave” to a USB port of your Linux PC and apply power to the 12V DC input.
The rpiboot tool should boot up the CM4 over USB and flash the new EEPROM image and boot order settings. Once the command line spew stops and the power LED on the dev board starts to blink rapidly, the process is complete and you can disconnect power.
Step 2: Try to boot from NVMe, fail
At this point I already had an ARM64 Ubuntu 20.04 LTS image downloaded from here and uncompressed to my Gigabyte GP-GSM2NE3128GNTD SSD. So I removed the SD card from my CM4 dev board and tried to boot from the NVMe.
I got a surprisingly informative error message from the Raspberry Pi bootloader (on a monitor connected to HDMI0)
start4.elf is not compatible. nvme boot requires newer software
So my first attempt at a fix was to update a 32-bit Raspbian SD card image with rpi-update and then copy the start4.elf and fixup4.dat files from that image to the NVMe boot partition. This got me to the next problem: The Ubuntu 20.04 image uses a version of U-Boot that doesn’t support NVMe, so it just sticks waiting for a boot device.
Next I tried downloading the Ubuntu Server 21.04 image and burning it to a SD card. I planned to go through the full process documented in the forum thread, starting with upgrading the Ubuntu image, so I inserted the SD card and powered up.
Step 3: Unexpected success
Imagine my surprise when it booted into Ubuntu 20.04 from the NVMe instead of 21.04 from the SD card as I was expecting. 😀
I guess what must have happened is that the boot partition on the 21.04 SD card had a version of UBoot that recognised the NVMe and treated it as a higher priority than the SD.
The next gambit would be to copy the boot partition from the SD card to the NVMe and see if it can boot directly from NVMe. But it works well enough as is, I don’t want to break it! This is left as an exercise for the reader 😉
Step 5: Try to make it boot directly from NVMe
Unfortunately on subsequent boots it flipflopped between NVMe and SD card as the root partition. So I thought since it’s Friday afternoon I should try to get it to boot directly from NVMe.
I started by deleting all files in the boot partition on the NVMe and replacing with the ones from the Ubuntu 21.04 SD card. The kernel is in /boot/vmlinuz, while the modules are stored on the root filesystem, so this would normally cause the modules to fail to load, except I noticed that I somehow managed to update the Ubuntu 20.04 installation on the NVMe so it was running the same kernel and modules as 21.04.
This still didn’t work, so I went back to that forum thread and replaced start4.elf and fixup4.dat with the ones from the updated Raspbian image.
Finally I got a direct boot from NVMe, with what I guess is Ubuntu 21.04 kernel and 20.04 userland. That works fine for me, I’d rather it ran a LTS version of Ubuntu.
We plugged it into the 3 phase outlet and it started up normally! 😀
When smashed with 750V the Tesla cabin heater would draw an impressive amount of power while warming up. Unfortunately the steady state power draw was only about 2kW, probably something to do with the rather weak fan cooling it.
The PFC line current waveforms at (roughly) this 11kW output power level. No surprises here, they look exactly like the theoretical ones for this circuit. (Except strictly speaking the red one is upside down 🙂 ) The theoretical power factor for this waveform is 0.95.
We don’t have a 3 phase power analyser in the lab, so I used 2 single phase ones on the input, according to the old “2 wattmeter method“. To be honest this didn’t work very well, as the power drawn by the PTC heater was always changing, and it was impossible to make sure the 2 meter readings corresponded to exactly the same time. Also, the PF reading is rubbish due to the inherent 30 degree phase shift: to get the actual PF you have to plug the wattage readings into a complicated formula.
When the heater reached steady state, I measured an input power of 2000W, an output of 1900W, and a power factor of 0.96. From an academic point of view it would have been nice to measure the efficiency at higher powers, I expect that 100W is mostly switching losses and the efficiency will increase with heavier loads.
The main goal was to get confidence that the PFC would work at its first gig, and this has been achieved 🙂
Another problem I have with the cabin heater testing (And the PFC testing 🙂 ) is that the PFC is not designed to start up into a load. This isn’t a problem for the intended application, as the Tesla coil won’t draw any current until it’s commanded to. But to power a resistive load, I have to use a switch to connect the load after the PFC has started up.
For the previous immersion heater tests I used an ordinary 240V AC rated switch. This will make a DC circuit, but won’t break it: the arc simply won’t go out until the entire switch is incinerated. Of course if you’re reading this there’s a fair chance that you actually enjoy burning things to a crisp with electric arcs, but I have to be on my best behaviour to get the keys to the 3 phase outlet at work, and proper DC rated contactors are getting cheaper anyway thanks to the proliferation of electric vehicles.
The Kilovac LEV200 has hydrogen filling and a permanent magnet to blow out the arc. It’s good for several hundred amps at 900V DC.
I thought I’d better test the characteristics of the PTC elements in a scientific manner as opposed to just applying 750V DC from the PFC and seeing what happens.
When measured with a multimeter at room temperature, each element read about 600 ohms. This would imply a rather low power output if that was the minimum resistance.
For my next experiment I connected one of the elements to my old Xantrex 600V bench power supply. The 600 ohm cold resistance would imply a draw of no more than 1A at 600V if the element was purely PTC. But to my surprise, the current draw actually began to increase as the element got hotter, eventually hitting the PSU’s 1.7A current limit at only 200V.
This means that the elements must actually start off as NTC, and transition to PTC at a higher temperature. That kind of makes sense, as a car heater matrix has to be able to start up from very low temperatures. Purely PTC elements would presumably exhibit an immense current surge when the heater is turned on after leaving the car parked overnight in Canada. 🙂
I was wondering why Tesla bothered to implement individual control of the 6 elements, and I guess the turn-on surge is the answer to this too: by turning them on sequentially the surge can be made 6x smaller. It’s not like a Tesla traction battery (or even an IGBT) would care if the turn-on surge was 10A or 60, but maybe it allows them to use a lower rated fuse to connect the heater to the battery. Fuses rated for high voltage DC are expensive so the savings made here might outweigh the cost of 5 IGBTs and drivers.
Anyway, I couldn’t test above 200V due to the limited output current. The next larger PSU I have is the PFC, and it has a minimum output voltage of 400V (600 on 3 phase) so I will just have to “send it” as the kids say nowadays.
I needed a high power dummy load that was a bit more health and safety friendly than my bucket of immersion heaters. I investigated lots of possibilities until I eventually found a Tesla Model 3 heater matrix on eBay. The Internet said it could draw up to 6kW, and being PTC, it shouldn’t catch fire if I forget to turn the fan on, which would look great on a risk assessment. So I went for it. 🙂
The lid is held on by penta-lobe screws with a tamper-proof peg in the middle. There was nothing in my collection of tamper-proof bits that would fit, but a sturdy flat-bladed screwdriver worked quite well after Dremeling a new slot or just jamming it in hard enough to break off the peg. 😀
Ooh, fancy! What does all this stuff even do?
There appear to be 6 separate PTC elements, each with low side switching by a 600V IGBT and non-isolated gate driver.
To the left of the IGBTs is a voltage divider and low side current shunt, and on the right an isolated CAN interface and DC-DC converter. Handling CAN communications and A/D conversion of the voltage and current signals, we have an 8 bit ST microcontroller (probably sharing its 0V rail with DC bus negative)
Of course I immediately set to work reverse engineering the CAN protocol so I could command it to connect its elements in 2 groups of 3 for 800V input. Oh wait that’s not gonna work 🙁 we need a hardware solution…
The heating elements were connected to the PCB by spring clips that released quite easily, but the PCB was stuck firmly to the aluminium enclosure with thermally conductive glue. I freed it with the old embedded programming trick of heating the enclosure with a heat gun and prying with a paint scraper. (sadly this doesn’t work any more in Python 3 😉 )
With the PCB removed we can see the terminals for the heating elements, looks as if some sort of spade terminal should fit nicely. (Also note the little nodule at bottom right which appears to be for cooling the CAN transceiver chip.)
And after a pulsating second half the score is Wago 221, CANbus 0. 😀
Since Carbon Intensity didn’t offer a handy regional forecast widget, and JSON APIs are almost fun, the obvious course of action was to grab a Raspberry Pi and make something to inform my electricity consumption decisions.
I found a Sense HAT in my drawer of IoT junk (some might say it’s more like an entire building than a drawer) so I used the RGB LED matrix on that to display the result. It fades from green at an intensity of 0, through to red at 215g CO2/kWh, which is approximately the carbon footprint of natural gas burnt for heating. All LEDs are programmed to the same colour, and it is covered with a globe from a broken LED light bulb to make it look like a single light source.
Carbon Intensity’s forecasts are updated every half hour, so I pull the 24 hour regional forecast from their API a few minutes after each half hour, and crunch it down to a single number representing the average carbon intensity for the next 3 hours.
I got it to work and amplify, but the loop gain left a lot to be desired, so I decided to start over on the compensation. I also hooked up an unregulated power supply and a different output stage, partly because I wanted to see how the PSRR was doing, and partly because I wanted to reassemble the Ice Block with its original driver board.
Now, you should never anthropomorphise amplifiers, they hate it! I swear that this one “wants” to blast electronic music from the 90s at high volume though. 🙂
Having given the Ice Block its output stage back, I had to find another one for my experiments. A search of the junk pile yielded the remains of a Maplin 100W MOSFET amp kit. I’d have preferred to try BJTs, but the PG508 prototype was already set up to work with lateral MOSFETs.
I tried the time-honoured method of soldering RC networks in random places, or maybe places that seemed to make a difference when touched with a damp finger. 🙂 This improved it somewhat, but it still wasn’t doing a great job of correcting the output stage’s copious (and vintage correct!) crossover distortion.
I eventually cracked Bob Cordell’s “Designing Audio Power Amplifiers” and spent an afternoon pondering Chapter 9, “Advanced Forms of Feedback Compensation”. It struck me that the PG508 topology is very similar to figure 9.7, except that the input stage doesn’t exist as such: the feedback node is the VAS input.
It also struck me that I’d already ended up with RC networks in the places shown in fig. 9.7 by trial and error, just with completely different values. R4 and C2 were in the original Tektronix PG508, and R5 C3, R3 C1 were my additions. So the obvious course of action was to change them to the values suggested by Cordell and see what happened.
Initial results weren’t great: it oscillated at 20MHz, but this was squelched by reducing R4 to 51 ohms. Having done this, performance was excellent: the 16pF C1 gave the extra loop gain I was looking for. I’d started out with 100pF here as that’s the value used in a Douglas Self Blameless amp. The Blameless input stage typically has 5-10x the gm of the PG508’s non-existent IPS, though. So funnily enough C1 needs to be 5-10x smaller to get comparable loop gain.
With these modifications the measured performance was 0.03% THD at full power at 10kHz, and 0.00something at 1kHz. The 10kHz figure seems high, but it’s now in the ballpark for a well functioning driver doing its best with a vintage MOSFET output stage. (Cordell’s AES paper quoted 0.02% at 10kHz with the Hawksford error correction turned off.)
Note that this THD figure is no better than I got with the old compensation and the Ice Block output stage. This just means that the Ice Block output stage must have about 3-5x less distortion than the single 2SK135/2SJ50 pair used here.
Slew rate was also improved, and stability with a capacitive load was just about acceptable: with 0.1uF slapped on the output it showed a few cycles of damped ringing but didn’t oscillate.
I also took the opportunity to test out the opamp front end inspired by the Quad 405 and Cordell Super Gain Clone. I used an OPA2604 as it was the best opamp I had around. This works very nicely: it reduces the DC offset to 2mV, undoes the phase inversion inherent in the PG508 circuit, and increases the overall gain from 10 to 50.
Note that the opamp must be a FET input type because of the high impedance of the DC feedback path. Also, as the circuit has 2 LF time constants (the 1M/1u and the 47k/2u) with feedback around them, it functions as a 2nd order active high pass filter. It rolls off at 12dB/octave and can resonate if the time constants are too close together.
I basically copied this part from the Quad 405, so it must have done the same thing. I guess it was desirable to have a good rumble filter here in the days of vinyl. Arguably it still is in the era of small vented speaker cabinets and dubstep. 🙂
I was so excited (honestly 8) ) about the idea of a PG508-based audio amp that I decided to try building it in real life.
I first tested the driver circuit with +-15V supplies, no output stage and pure dominant pole compensation. It oscillated happily at 6MHz, and to get it stable I had to go back to the original PG508 lag compensation network (the 51 ohm and 8n2) The back-to-back diodes helped the clipping behaviour: without them you can reverse bias the cascode transistors and burn out the LEDs if you really overdrive it.
In the interest of getting something working quickly, I hooked it up to the output stage and power supplies of my ancient and long-suffering Ice Block amp, which conveniently happened to be partly dismantled with one driver board missing. It provides regulated +-65V and +-15V rails in addition to the main +-53V, so no worries about PSRR for the time being.
To achieve stability with the Ice Block’s hefty lateral MOSFET output stage in the loop (2 pairs of double die Exicon FETs) I had to use both the original PG508 compensation and dominant pole compensation with an extra zero (the 1k and 100pF).
I think it’s a bit temperamental because there are 3 transistor stages in the loop enclosed by the dominant pole capacitor. The Douglas Self Blameless only has 2, and my previous attempts at adding a third stage to that (cascode connection of Cdom) also caused oscillations at a few MHz.
As with any half decent solid-state amp, THD+N was at the limit of my measurement system at 1kHz, and dominated by noise. I had to go to full power at 10kHz to see a meaningful distortion residual.
The distortion appears to be what Douglas Self called “gm-doubling”: in a push-pull circuit the gain is higher when both halves of the circuit are contributing, than when one half is cut off. And in the residual we see small lumps corresponding to increased gain around the zero crossings. I’m not entirely sure what part of the circuit is causing it. It may be the output stage, as that’s the usual culprit. It looks like it wants less bias, but the bias pot won’t go down any further. (Got to replace that TL431 with a TLV431)
I also tested the full power bandwidth, and it happily delivers 100W to 100kHz and beyond. I didn’t push it beyond 130kHz for fear of burning out the Zobel network.
A reading of 0.025% at 100W and 10kHz, with no filters engaged, is not to be sneezed at. I’d be perfectly happy with it, except the other un-hacked channel of the Ice Block does 0.009%! The PG508 circuit has some way to go before it can beat the original Alexander CFB.
While I had the equipment out, I also tried measuring distortion with a LF411 in place of the SSM2131 in the Alexander circuit. It made no noticeable difference at any power level, even though I’d persuaded myself that the LF411 sounded bad…