How far to put it
- About 60 mm from the card's back (55–65 mm), square to it, focused on the board. There it sees 40 × 30 mm, the whole die with about 4 mm to spare above and below, at 0.16 mm per pixel. Focus holds while the board stays within ±2.2 mm of where you focused (±4.4 mm if 2 pixels of blur is fine), and 12 mm stay free behind the camera. The distance is to the lens's optical centre, which may sit a few millimetres behind the camera's front face, so the front may end up at 50–60 mm. You focus by eye anyway.
- Closer isn't sharper in practice. Each millimetre closer makes the pixels finer but the zone in focus thinner. If the arm holds the camera to ±2 mm, the finest detail you can count on is about 0.17 mm at 44 mm. Anywhere from 33 to 59 mm is within 10% of that. Nearer than 44 mm, the blur from a 2 mm sag grows faster than the pixels shrink.
- Farther is capped by the gap. Past about 72 mm the camera's back hits the cooler. The die fits whole from 46 mm. The package (45 mm) fits across from 67 mm, but top to bottom only from 88 mm, so never whole inside the gap.
- For heat pictures the distance barely matters. The board spreads the chip's heat over 12–48 mm, which is dozens to hundreds of pixels at every distance here. Between 45 and 70 mm a shire-sized patch (3.7 mm) averages 400 to 1,000 pixels, so its noise is 1.3–2.0 mK per frame (your unit, raw: 41 mK per pixel). The distance changes how sharp the parts look, which decides how well you can tell real heat from reflections.
- The focal length is fixed at 4.3 mm. What you choose is where the camera sits and where you turn the focus ring. The ring has no encoder, so focus by eye on the live view: turn it until the capacitors' cold dots are sharpest, then tape the ring.
The simulator
Drag the camera in the top or side view (the orange knob behind it tilts it), or drag the green outline on the card's back to aim. The picture is what the P3 would show: the chip's heat as a broad glow, the capacitors as cold dots, blurred by the focus. Everything is to scale.
What the P3 sees
256 × 192 pixels, drawn as the camera's own pixels.
The card's back, from the cooler's side
The green outline is the camera's view: brighter where it's in focus, red where it isn't. Drag it to aim. The rear wall is on the left.
From above
Drag the camera to move it, the orange knob to turn it, the dashed line to move the focus. The cone is the field of view; the shaded band is the depth that stays sharp.
From the side
Looking along the card from its tail. Drag the camera up and down; the orange knob tilts it.
Sharper pixels against thinner focus
The smallest detail you can count on, if the camera stays focused on the board give or take the drift you set below. Pixels and diffraction grow with distance; focus drift shrinks with it. Drag on the chart to set the distance.
Settings: focus, drift, tilt, orientation, board
The three limits
- Pixels. A 12 µm pixel through a 4.3 mm lens covers 2.79 milliradians, a little less up close: 0.16 mm of board at 60 mm, 0.07 mm at 30 mm.
- Diffraction. At the camera's 8–14 µm wavelengths an f/1.0 lens can't focus a point smaller than about 24 µm, which is two pixels across (the Airy disk at 10 µm). The lens and the pixels are matched, so focus better than about one pixel of blur gains nothing.
- Focus. The ring sets one distance that's sharp. The zone that stays within a pixel of blur grows with the square of the distance: ±0.5 mm at 30 mm, ±2.2 mm at 60 mm, ±5 mm at 90 mm. A gooseneck that sags, a board that warps or a tilt across the view all eat into it: a 3° tilt across 40 mm is 2 mm of depth.
The chart puts the three together for a mount that holds focus to a given ±. The best distance moves out as the mount gets sloppier: about 31 mm at ±1 mm, 44 mm at ±2, 54 mm at ±3 and 70 mm at ±5. At every setting the curve is flat near its bottom, so the field of view and the gap decide the rest.
| Lens to board | Field on the board | Per pixel | Sharp within (1 px / 2 px blur) | Finest detail, mount ±2 mm | What's in view | Room behind the camera |
|---|---|---|---|---|---|---|
| 30 mm | 18 × 14 mm | 0.07 mm | ±0.5 / ±1.0 mm | 0.20 mm | the middle of the die | 42 mm |
| 45 mm | 29 × 22 mm | 0.11 mm | ±1.2 / ±2.4 mm | 0.17 mm | the die's width, almost its height | 27 mm |
| 52 mm (bracket in slot 2) | 34 × 26 mm | 0.13 mm | ±1.6 / ±3.2 mm | 0.18 mm | the whole die | 20 mm |
| 60 mm | 40 × 30 mm | 0.16 mm | ±2.2 / ±4.4 mm | 0.19 mm | the die and the capacitors around it | 12 mm |
| 65 mm | 43 × 33 mm | 0.17 mm | ±2.6 / ±5.2 mm | 0.20 mm | most of the package's width | 7 mm |
| 70 mm | 47 × 35 mm | 0.18 mm | ±3.0 / ±6.0 mm | 0.21 mm | the package's full width | 2 mm |
| 90 mm (needs a mirror) | 61 × 46 mm | 0.24 mm | ±5.0 / ±10 mm | 0.26 mm | the whole package | doesn't fit |
Checked against your camera's published spec
Rechecked Thu 15 Oct 2026 (week 42), 00:20 PDT, against Reverse-engineered Thermal Master P3 (your unit, serial P3025043DG040200256), with every number recomputed. The answer holds: about 60 mm from the lens's optical centre, focused on the board. No input the spec allows moves any result by more than 10%.
| Input | Used here | Your camera, per the spec page | What it changes |
|---|---|---|---|
| Focal length | 4.3 mm | 4.3 mm, from the serial number's lens code "043" (tenths of a millimetre). The page's own 40.0° × 30.2° field implies 4.22–4.27 mm, and its 0.160° per pixel would make 40.9° across. | ±2% on the field and the pixel size: 39.8–40.6 × 29.8–30.5 mm at 60 mm. Nothing else moves. |
| Aperture | f/1.0 | f/1.0, from the spec sheet. The official app writes f/1.1 into its photos, a bug copied from the P2. | At f/1.1, sharp within ±2.4 mm instead of ±2.2, a 27 µm diffraction spot instead of 24, and the best distance 41 mm instead of 44. |
| Sensor | 256 × 192, 12 µm | Raytron RTC2121W: 256 × 192 at 12 µm, 8–14 µm, 25 Hz. Confirmed by the frames themselves. | Nothing. |
| Focus range | 8 mm to ∞ | The same. It's a mechanical ring with no encoder, and the 8 mm is the vendor's claim. | Untested. The page's "35–50 µm sampling at 8 mm" fits only if the 8 mm is measured from the front of the lens and the optical centre sits 9–14 mm further back. A thin 4.3 mm lens 8 mm away would sample every 10 µm. |
| Where distance is measured from | the lens's optical centre | Not documented. | If the centre sits some millimetres behind the front face, the front is that much nearer the board, with that much more room behind: try it under Settings. On Friday, a ruler in the view at a measured distance gives the real number. |
| Body | 59 × 27 × 17.2 mm | 59 × 27 × 17.2 mm, 26.3 g. Your measure on Friday read 59 × 28. | Nothing. |
| Noise per pixel | 41 mK per frame | 33.8 mK (median) with the camera's own temporal filter on. Raw, in the high-precision mode, the noise is 1.07 steps of 38.62 mK, so 41 mK. | The filtered frames aren't independent, so the raw figure is the one to average. A shire-sized patch at 60 mm: 1.7 mK per frame, or 0.35 mK over a second once drift is removed. The page used 34 mK until this check. |
What else is assumed
- A thin lens. A real 4.3 mm lens with several elements behaves the same at these distances apart from where its optical centre sits, which is covered above.
- The lens sits in the middle of the 59 × 27 mm face. Nobody has documented where it is.
- The chip's centre is Friday's estimate: 86 mm in from the rear wall, 42 mm up on a low-profile board and 57 mm on a full-height one. The capacitor field behind it is drawn from the 15:35 live view. The screws and the power stages are illustrative.
- The picture's temperatures are illustrative (about 44 °C board, 52 °C over the die, shiny parts 6 °C cold). The blur is computed per pixel from its depth, with diffraction at 10 µm.
- "Finest detail" is the width of the blurred spot on the board: the pixel, diffraction and the defocus at the worse end of the drift, added in quadrature.
Sources
- The camera: Reverse-engineered Thermal Master P3, published by your Antigravity session, and its source file
THERMAL_MASTER_P3.md(~/git/thermal-androidon the M5): §1.2 the 256 × 192 Raytron RTC2121W at 12 µm and 2.79 mrad per pixel; §1.3 the noise of your unit (33.8 mK frame to frame, median); §1.4 the 4.3 mm lens from the serial number, the 8 mm–infinity ring with no encoder, f/1.0; §2 the 59 × 27 × 17.2 mm body. - The rig: Friday thermal measurements (the 89 mm gap, the card's 168 mm, slot 5 of 7, the low-profile board, the chip's place) and the thermal camera hub (the board blurs heat over 12–48 mm).
- The model: thin-lens defocus (blur = aperture × the image-distance error ÷ image distance), Airy disk 2.44 λN, and depth of field from the same formula. The page's code is its own record (view the source); the recheck's script is
~/git0/aifoundry/work/friday-thermal-measurements/scripts/optics2.pyon the Intel.
Total life satisfaction: the thermal model is one of the three goals at the top of your Project doc. With the camera's real lens and Friday's measurements, "how far?" has an answer you can check by dragging: about 60 mm, focused on the board, with the ring taped.