ATX PSUs have poor regulation on the 12V rail but this doesn't affect the rest of a 3D printer because everything is regulated downstream. The stepper drivers are constant current and the heaters are temperature controlled.
Here is a dip in the 12V rail caused by the 10A bed switching on. As you can see the swing is 12.2 to 11.6V, around 5%. This results in about a 20% reduction in light which is very noticeable.
One way to fix the problem would be to make an accurate 12V from the unused 5V rail using a small boost converter. I didn't want to cause any EMC issues, so I decided to use a low drop out (LDO) regulator to regulate to just below 11.6V. I thought it would just be a single chip solution, but after a lot of searching, the lowest drop LDOs I could find were 0.5V and I didn't want to waste that much.
I decide to roll my own using a low RDSon MOSFET as the series element. The LEDs only take about 0.5A, so the drop out voltage with a MOSFET with an RDSon of a few milliohms can potentially be a few millivolts, hundreds of times better than any LDO chip I could find.
In practice I used a BTS134 MOSFET I had lying around with an RDSon of 60mΩ. To put that in perspective that is less than the voltage drop in the wires I used.
This is the circuit I came up with: -
I use a TL431 adjustable precision reference as the control element. These are great little chips that can do anything from replace a zener diode to being an audio amplifier. The green LED acts as a level shifter to ensure the cathode voltage of D1 is higher than the reference voltage, a requirement of the chip. A white LED would give more margin.
VR1 should be initially set for the maximum output voltage and then gradually reduced until the flickering stops when the bed is switching. Handily the green LED flickers when the LEDs are flickering aiding adjustment without being blinded by the light.
C1 prevents the circuit oscillating at about 50kHz by providing negative feed back. I found its value by trial and error as there wasn't any phase shift versus frequency data in the data sheet. Basically R3 and R4 form an RC network with the gate capacitance of the MOSFET. This will produce a 90°phase lag at high frequencies. The TL431 must also create a 90°phase lag around 50kHz. To prevent oscillation C1 has to reduce the loop gain to less than one when the total phase shift is 180°. The circuit was stable with 1nF but had a bit of ringing. With 10nF it was a bit slow to respond.
The scope view below shows the performance of the regulator. The yellow trace is the 12V supply to the LEDs. The cyan trace is the negative supply to the LEDs (shown at a bigger scale). The purple trace is the difference, i.e. the voltage the LEDs see. It is virtually constant and the drop across the regulator goes as low as 40mV.
Here is a veroboard layout done in VeeCad: -