It runs · CPU hunt, part III

24 Oct 2026

523 joints

Saturday was soldering. JLC did the small parts on the back; everything you touch was mine: 73 pots, 18 jacks, 15 LEDs, four keys, the sockets for the Patch Submodule and the power header. I counted the joints afterwards, from the board file rather than the board: 523.

The back of the long green Rev A board, seen at a steep angle, lying face down on the black aluminium plate in a wooden case. Rows of fresh solder joints run across it among small black chips, an orange Patch Submodule sits plugged into its sockets near the far end, and a black shrouded power header stands near the front edge. A reel of solder wire stands on the wood behind it. — the back, the Patch Submodule in its sockets

The front of the green Rev A board on a pale wooden box, seen low along its length. Rows of bare metal pot shafts on green pot bodies fill it, red LEDs on long legs stand between them, a row of black jacks runs along the left edge and a red key sits in the front corner. — the front, before the plate

The black aluminium plate on top of the board, every pot shaft poking through its hole, with strips of yellow masking tape stuck over the plate in a dozen places. A reel of solder wire stands behind it on the wooden box. — plate on, tape on, flip it over

Rev A is done except the SD card slot. It is still sold out everywhere, but three are on their way as free samples and get soldered in next week. The board runs without one anyway.

First boot: the firmware flashed, all 73 pots read, every lamp lights. The four keys read pressed with nobody touching them. I had soldered them in the wrong way round, so they are closed at rest and open when pressed. Desoldering four keys from a finished board is no fun, so the firmware got a switch that reads them the other way round.

Then a click: switching a deck to the bucket brigade made one audio block take seven times its budget, because it cleared two 1 MB tape lines in one go. Now the lines only forget what was written, and ten minutes hands off brought no click.

A close-up of the board's back: the orange Patch Submodule sits in its sockets on the green board, surrounded by small black chips, solder joints and white part labels. Below it, two small empty pads marked C_SD1 and C_SD2 wait for the SD slot. — the empty pads below wait for the SD slot

The finished module in a wooden-sided eurorack case on a table by a window, powered. The black plate holds rows of bare silver pot shafts, a few of them glowing red where LEDs sit underneath, red dots of light scattered across the fields, a green key at each end and a red one on the right. A power cable and a purple ribbon lead in at the right end. — first light, no knobs yet

The same case in daylight with the knobs on: green caps on the left, dark blue ones in the centre column and orange ones on the right, in even rows above a line of jacks. A few patch cables hang from the right-hand jacks, and a bonsai and a potted plant stand on the window sill behind. — and dressed

And the Thonk parcel's bill finally came: 81.75 € duty and import VAT. The hardware ends at about 700 €.

25 Oct 2026

Noon

On the first boot every pot sat at its left stop. Turn them all to noon and the patch wakes up, and so does the old problem: the audio callback has 945 000 cycles per block, and at noon it took up to 1 030 000. It crackled.

The finished module from straight above, lying in its open case on a wooden table, powered. Green knobs fill the left half, dark blue ones the centre column and orange ones the right half, most of their pointers standing near the top, with a few red LEDs lit among them and a green and a red key at the ends. A row of jacks runs along the bottom edge, two cables lead out of the right-hand outputs, and a power supply module sits in the case at the far right. — most knobs near noon

Yesterday evening and today went into finding where. TIDE searched its whole step grid every time it moved a lane; now it checks the step it is on first, and its worst block fell by almost half. The wavetable voices paid double for their mip crossfade while struck; a cheaper crossfade gives the same bits.

The noon patch now peaks at 875 374, under the 900 000 I want as margin. Two releases, 2.23.2 and 2.23.3, CPU work only. Nothing sounds different, which is the point.

27 Oct 2026

Seven quiet ones

RANGE on a note deck used to push the melody up its span, so at half-way a tune parked near the top and fell onto the root. Now RANGE is a width around a fixed centre: turn it down and the melody gets narrower, not higher.

Then a hunt for bugs you only hear once you know them. Seven. The best one: BODY applied velocity twice, so a soft accent of 0.3 landed at 0.09, more mouse than drum. 2.23.5 fixes all seven, and TUNE fully left now goes an octave lower, down to 55 Hz.

28 Oct 2026

Where you left it

Until today every power cut was a factory reset. Now the board remembers. The knobs speak for themselves at boot; the flash keeps what they can't show, like the 50 depths under MOD. It saves once the knobs have been still for 15 seconds, at most every five minutes.

off at two o'clock, on at two o'clock, and the flash did the remembering

The first try played the old filter for 30 ms after boot, so the output now stays mute until the knobs are read, then fades in. And one group of pots jitters by up to 57 counts on its own, which the board took for a hand still turning, so it never got to save. Those pots got a wider dead band.

Also new: SHIFT + MOD wipes the MOD layer, and its lamp blinks five times to say so. The ENG knob now runs the note engines first, then BBD, then Sampler.

The first five minutes recorded off the hardware rather than out of Rack, straight from the line out:

Session recording · Rev A · the first take off the hardware

0:00

--:--