something I've wanted to build for a long time is one of those big LED cube matrices, but with precision temperature sensors on each vertex so that the pattern can show the differential temperature across the cube.

the problem is that I haven't come up with a way to make it anything close to affordable. on an 8x8x8 cube you need 512 sensors.

1-wire would be perfect but it's way too expensive (mostly due to patents).

precision I2C temp sensors are pretty cheap now, but even ±0.2°C precision (e.g. STS40, ~£205 total cost for sensors) is too much error for measuring thermal differentials across a cube, and you end up needing a lot of I2C buses due to addressing limitations.

placing an NTC on each vertex is an attractive option due to the low cost per unit and the precision being limited only by your acquisition approach, but you still need to contend with acquiring 512 channels of data with sufficiently low noise and high resolution to capture the variations.

making the individual vertices autonomous isn't an option.

even 1% tolerance on the passives isn't good enough, and there's no good way to calibrate the variances out on a node-by-node basis.

as soon as you try to come up with a centralised calibration approach you're back to dealing with 512 analog signals.

it also really limits creative freedom.

ultimately I think the NTC approach is the way to go.

you figure out the optimal cost balance between analog muxes and ADCs, so your 512 channels get turned into banks of 32 or 16 final channels to be read.

then you use an opamp array to do a subtract and scale operation on those channels, so the voltages being fed into the ADCs allow you to get the most out of the full quantisation range.

calibration ends up being reasonably easy this way, too.

stick it in an insulated foam box, close the lid. hang the box on a wire so all sides have the same thermal losses. wait 10 mins for equalisation. spend a minute recording the temperatures. at each point in time, calculate the mean temperature and subtract it from the sensor values to get an error value. at the end you take the mean of each sensor's error as its calibration offset.

do the same in a few ambient temperatures and you can calculate the gradient variation too.

could everyone please stop replying before reading the existing replies, it's getting rather tiring replying to the same issues I've already addressed multiple times. thanks.

anyway, yes, NTC works out to be feasible cost-wise after all, at least at 8x8x8. the muxes are cheap.

power gets kinda fun because the peak current is 29.2A if you make all the LEDs full brightness white at the same time. naturally I will not be doing that. but even so, a 20A current budget seems reasonable. I will probably power it from 12V and run four 5A buck converters on the controller board. I've been meaning to design a supply with four 90° phase-shifted synchronised bucks anyway.