Thermal Master P3 on chips: what to watch and read — Tue 6 Oct 2026 (week 41)

Written Tue 6 Oct 2026 at about 18:00 on the Intel, for your ask: videos and docs on using the P3 to monitor chip temperature, and who else has done something like it. It adds to the 3 Oct camera page rather than repeating it.

Since then: every thermal page is gathered in the thermal camera hub. Where the camera fits behind aifoundry2's card, measured from your 9 Oct photos, is in Friday thermal measurements (Wed 14 Oct).

Start with these three

1. p3-ir-camera, and its LOCK-IN.md above all. This is the Python driver for your exact camera (USB 3474:45A2), and it now has lock-in thermography built in. Press l in the viewer and it shows four panels: in-phase, quadrature, amplitude and phase. With a 1 s period and 60 s of integration, it clearly showed 330 µW in a 33 kΩ resistor, which is below the camera's noise floor. It runs on Linux, and on Windows with a Zadig driver swap, but not on macOS. Last push: 7 Feb 2026. - The author is Joshua V. Dillon (GitHub jvdillon, Mountain View, an ML bio). He is probably the Josh Dillon you spent Sat 31 Jan 2026 (week 5) with, along with Christian Szegedy, Dumitru Erhan and Uliana Popov. That's not confirmed; it rests on the name and the Mountain View/ML match. If it's him, he may be the one person who has already run lock-in on a P3.
2. Dmytro Engineering's lock-in video: Hacking the T2S+ Before It Hacks Me – Free Lock-In Thermography Included!, with its write-up and code, IR-Py-Thermal (184★). Dima, in Ukraine, used a cheap Xinfrared T2S+, with a Raspberry Pi Pico switching power through a MOSFET. A 10 Hz blinking LED showed up after 10 s of integration. After 100 s he could see heat inside an RP2040 and a 1.6 mW voltage divider. He also imaged a decapped AD9042 die. Hackaday covered it (28 Oct 2024).
3. Christian Herglotz, Simon Grosche, Akarsh Bharadwaj and André Kaup (FAU Erlangen-Nürnberg), Component-wise Power Estimation of Electrical Devices Using Thermal Imaging (IEEE Trans. Consumer Electronics, 2021). They used a low-resolution consumer camera to estimate the power of each chip on a board (SoC, PMIC, Wi-Fi) with no special coating. The mean error was about 10%, but only for parts dissipating over 300 mW. It is the closest published thing to reading energy per chip off a cheap camera.

Videos about the P3 itself

I picked these by title and channel and checked those through YouTube; I haven't watched them.

The best written review: Gough Lui

Gough Lui, 7 Oct 2025 is the most technical review, and four of its findings matter for your rig:

• Turn X³ off for measurement. The 512×384 "super-resolution" makes small hot spots bloom, and their edges dance around. The sensor is 256×192.
• The phone and the PC save different files. A radiometric JPEG is 1440×1080 on Android and 512×384 on Windows, so you can't re-analyse one on the other. Raw frames through Josh Dillon's driver avoid the problem.
• The shutter is loud and fires on its own, which matches the 3 Oct page's "about every 90 s".
• The body is metal. Keep it off live boards.

Reading a chip's temperature correctly

• Tape it, then read the tape. FLIR's note Use Low-Cost Materials to Increase Target Emissivity gives Scotch 88 black vinyl tape an emissivity of 0.96. Shiny metal and bare silicon read wrong. At room temperature, bare silicon is partly transparent in the long-wave band, so the camera sees whatever is behind it. That's why the 3 Oct page put Kapton on the bare metal.

People who imaged running processors (the research lineage)

Every one of these groups had to remove the heatsink and replace it with something infrared-transparent. That's the line between their results and what a P3 can see on a cooled card.

• IBM, Hendrik Hamann et al., Hotspot-limited microprocessors: direct temperature and power distribution measurements (IEEE JSSC, 2007). They imaged a running PowerPC 970MP through mineral oil flowing between sapphire windows, then worked backwards to a power map per unit (they call it SIMP). The hotspot moved as the workload changed.
• UC Santa Cruz, Jose Renau's group. Power model validation through thermal measurements (ISCA 2007; the PDF is a scanned image), then Characterizing processor thermal behavior (ASPLOS 2010). They captured 125 Hz thermal maps at 10 µm pixels, with power logged at 1 kHz. ISCA's 50-year retrospective selected the 2007 paper.
• KIT, Hussam Amrouch and Jörg Henkel, Lucid Infrared Thermography of Thermally-Constrained Processors (ISLPED 2015). They argue that convection in the oil blurs the images. Instead they cool the chip from the back, which keeps a clear view of the die. They tested this on Intel 45 nm dual-core and 22 nm eight-core chips.

What it means for the ET-SoC-1 cards

• You won't see the die. You'll see the heatsink's top plate and the back of the card, both a few kelvin cooler than the die and seconds to minutes behind it. That is the 3 Oct plan's view, and these papers explain why it has to be.
• Lock-in is the right tool for that view, and on a card the stimulus is software. Run the workload on and off with a fixed period instead of switching a MOSFET. Josh Dillon's code sends 1/0 over a serial port. Point that at whatever starts and stops the kernel, and log the switch times.
• Herglotz's 300 mW floor is a single-frame number. Lock-in buys orders of magnitude: the two hobby set-ups above saw 1.6 mW and 330 µW.

What I didn't find

• No one using a P3, or any consumer camera, on an AI accelerator card. Your experiment would be the first written up.
• No open-source Android code for the P3's USB protocol, in two searches. Your Android app would be the first. The protocol itself is written up in Josh Dillon's P3_PROTOCOL.md.
• No video by a well-known hardware channel imaging a delidded GPU or CPU with a phone camera. The search turned up only Guru3D's FLIR board shots.

Earlier pages on this

• Thermal camera through a Pixel, Sat 3 Oct 2026 (week 40): the P3 rig, the mount and the streaming.
• The thermal-camera experiment plan, Sun 27 Sep 2026 (week 39): where to aim it, the lock-in sensitivity, 20 experiments.

Total life satisfaction: the thermal model is one of the three goals in your nekko.ai Project doc. Before you write code, the cheapest step is one message to Josh Dillon, if he's the Josh Dillon you know.