Cooling the ET-SoC-1 from the back, with the camera on top

AI Foundry. Written Fri 9 Oct 2026 (week 41), about 19:40–20:45 PDT, on the Intel (~/git0/aifoundry), on your ask: research mounting the camera with the chip cooled from the back. Your idea: a cable so the PCIe card can stand alone, a cooling solution on its back, and the camera on top. Built from the lab's measured heat law (fitted 7 Oct, checked against today's fan swap), the 30 Sep and 6 Oct heatsink pages, the physics model behind the thermal camera hub, and new research on coolers, pads, cables and stands. No card was touched.

The short answer

• It can work, and one unmeasured number decides how far. That number is how easily heat leaves the package into the board (its junction-to-board resistance, never published for this chip). If it is near 1 °C/W, a cooler on the back plus an air jet across the lid runs the card at 600 MHz with work on it. If it is near 3 °C/W, only the lowest clock idles safely.
• The first step is safe and settles it. Add the rear cooler with the heatsink still on and read how far the idle temperature drops. After today's new fans, expect 3 °C (poor) to 6 °C (good), not the 15–20 °C the 6 Oct page set as its bar: those thresholds were for the old, weaker cooling.
• The view from the top is far better than from the back. The lid carries 10 to 100 times more of the fluctuating signal, and its copper blurs heat over millimetres instead of centimetres. Single shires come within reach with lock-in.

• It needs a card of its own. The heatsink comes off and the card runs bare, so this is the card Roman offered to sacrifice, never a shared lab card.

• It has been done. Amrouch and Henkel at KIT (ISLPED 2015) cooled desktop processors from the back of the board, with a water block, a small Peltier as trim at 0.9 W, and a milled metal plate on the PCB. Their camera looked straight down at the die, and the steady temperatures matched the stock heatsink's. The rig also settled faster, with no heatsink mass over the die. Their chips ran tens of watts, the same class as this card.

• The build, about $350 with air or $450 with water: a powered cable kit, a laser-cut plate with a printed cradle, a cooler on the board's back over a soft pad, a 50 mm blower across the lid (it can run off the card's own fan header), a cutoff that latches and kills the host, and the P3 on a column 100–150 mm above. The build has the parts.
• The card's drawing settles a 6 Oct question. Its 6-pin aux 12 V joins after the card's current sensor and its protection switch, so power fed through it is invisible to the card's own power reading. Confirm that with a meter.

How much cooling a bare card needs

The lab's heat law for aifoundry2's card (board power climbs with die temperature, doubling its leaky part every 22.5 °C) reproduces today's idle exactly: 51 °C at 0.68 °C/W in the rig's 35 °C air. The same law gives the most thermal resistance each case can tolerate before it runs away:

A bench helps. Room air at 22–25 °C instead of the rig's 35 °C buys about a quarter more headroom. The card must survive the first row in any case: it idles at 600 MHz through the host's 30–90 s boot, before anything can lower its clock.

What the build reaches. Heat leaves two ways at once: down through the package into the board and the rear cooler, and up from the bare lid into the air jet. The two add like parallel resistors:

Read the two tables together: - At 1 °C/W into the board: idle and a working quadrant at 600 MHz are safe on a bench; full load only with a cold plate and a strong jet. - At 2 °C/W: idle at 600 MHz is safe; a working quadrant is marginal. - At 3 °C/W: idle at 600 MHz is marginal, the lowest clock is safe, and nothing heavier runs.

The tables mix two kinds of watts. The heat law counts the whole board's power, while the cooling estimates count only the chip's own heat, perhaps three quarters of it. So the totals lean pessimistic.

What the camera sees from the top

• Ten to a hundred times more signal than through the board. The fluctuating part of the heat reaches the top of a coated lid almost undiminished at 0.1–0.5 Hz (0.71–0.86 of it), against 0.005–0.09 through the board.
• The copper lid still spreads heat sideways, over about 8.5 mm at 0.5 Hz and 2.7 mm at 5 Hz. So quadrants show at once, and a single shire (3.7 mm) needs lock-in at a few hertz.
• The 30 Sep study's numbers for a coated lid: lock-in shows coarse patterns within about 10 s at 600 MHz; at 100 MHz, a single shire within about 10 s and a checkerboard within half a minute.
• One frame will look nearly uniform. With the heat going down into the board, the lid carries little steady heat out of its top, so its surface sits close to one temperature. The pattern lives in how it changes, which is what lock-in reads.
• Placement. A camera 100–150 mm above sees the whole 45 mm lid (a field of 68–104 mm, at 0.27–0.41 mm per pixel) with deep focus. Coat the lid: matte black paint keeps 94% of the signal at a 2 s period, 1-mil Kapton 79%. Bare plating reads the room. 1-mil Kapton only reaches an emissivity of about 0.62 over metal (corrected Wed 14 Oct; calibrating the P3), so paint is the better coat.

The build

The parts (prices approximate):

How it goes together: - The card lies flat on the cradle, components up. The cooler sits under the board's back, bolted to the plate, with hard standoffs setting the gap. It must not hang off the board. - Force, not the pad, is what to get right. The pad's own resistance is small: 0.33 °C/W at 2 mm and 3 W/mK over the package, 0.08 with the putty. But squeezing that pad 30% takes about 110 N over the package's shadow, enough to crack solder balls. Aim for 10–20% compression, about 2–4 kgf over four spring screws. - The blower across the lid runs steady, never PWM-modulated. A moving air jet sets the noise floor in thermography before the camera's own noise does. Record a fan-on, fan-off null before any lock-in run. - Coat the lid and calibrate it. The KIT group measured bare packaging at an emissivity near 0.01, which reads mostly reflection. Paint or Kapton the lid, then check the reading against the card's own thermal diode (the card has one, read through its power controller) at two power levels, with a patch of known emissivity in the frame. - The cutoff must latch and kill the host, not the card. A cheap bimetal thermostat (KSD301 type) resets itself about 15 °C below its trip point, so on a bare card that reheats in 1–3 minutes it would cycle on and off. A thermocouple controller with a latching relay on the host's power-on line or its mains avoids yanking a live device out from under the host. A one-shot thermal fuse is the last resort. - Skip the Peltier. Driven hard, it adds about 60 W of its own heat for the cooler to remove, and below the dew point it condenses water onto a live board. The KIT rig used one only as a 0.9 W trim on a water block.

Measure on site first (stage 0). The card's drawing (in AI Foundry's et-man repo) doesn't show how the heatsink attaches, and no photo online does either. With the heatsink still on, check: - whether four holes frame the package, and their spacing (58, 61, 65 or 70 mm are common) and diameter; - whether the heatsink's screws run into a backplate on the board's back, which would have to come off; - the tallest part behind the package, and the clear area around it; - the board's thickness, and the lid's height and flatness; - with a meter, where the 6-pin aux input joins the 12 V rail.

What the 9 Oct photos already show (Friday thermal measurements, Wed 14 Oct): there is no backplate, and the back is bare board covered in small parts. Two brass screw heads show on it, probably the heatsink's; their spacing is still to measure. The board also looks low profile, about 69 mm tall, under a heatsink and shroud that stand about 100 mm.

For the camera on top: at 100–150 mm the P3 frames the whole lid, at 0.27–0.41 mm per pixel; the lid's copper blur is coarser than that. Check the framing before buying the arm.

The plan, in stages

1. Stage 1, heatsink on, in the rig (safe; about an hour of card time). Fit the rear cooler to the board's back over a soft pad, without removing anything. Then read the idle temperature against today's 51 °C, averaged over ten minutes at 10 readings a second:

(Predicted with the heat law, for a cold plate and pad of about 0.35 °C/W; with a finned sink each drop reads about 1 °C less. Today's fans make every drop smaller than the 6 Oct page assumed.) 2. Stage 2, the card on its cable, heatsink still on. Move the card to the bench on the powered cable and check that the link still trains at 16 GT/s x8, the error counters stay flat, and the power reading is whole. Take a baseline with the camera on the heatsink's side panel. 3. Stage 3, heatsink off, the cutoff armed. Someone on site, a thermocouple on the lid feeding an independent cutoff at 80 °C, the air jet running, the rear cooler running. Power on, set the lowest clock at once, and watch for ten minutes. Only then try 600 MHz idle. 4. Stage 4, the pictures. Coat the lid, then lock-in on a quadrant, its opposite, and single shires, compared against the chip's own sensors.

Risks, and who says yes

• Roman Shaposhnik: the card itself (his 29 Sep offer named none), taking its heatsink off, anything clamped to it, the cable, and a host to run it in.
• The boot window. A bare card idles at 600 MHz until software lowers its clock. A firmware that boots into the 300 MHz safe state would remove that risk; it's a question for Roman.
• Nothing on the card stops a runaway. The 90 °C stop is the runners' policy, not a hardware limit, and aifoundry2's card has climbed to 138 °C with no work on it. The independent cutoff is not optional.
• Re-pasting changes the card. Once its heatsink comes off and goes back on, the card's fitted thermal constants are void.
• Power through a cable. A plain riser carries the slot's 12 V, rated 66 W, against the card's 88 W maximum. Use a cable whose card end takes its own 12 V, and check first whether the card's aux input passes its current sense.

Total life satisfaction: this is the route to the picture you described, heat moving shire by shire, and its first step is an hour of work with nothing taken apart.