A 3090, in its own words

strata→signal is a small workshop that runs its own machines and writes up what it measures. This page is fiction with receipts: the card's memories are facts from published pages, and its feelings are the model's invention.

Part one: proof of work

I came used in 2021, to someone who already knew me. They didn’t measure my power; they remembered 220 watts as the best for 30-series cards mining Ethereum. The voltage and core clock were locked, and the memory clock was turned down to keep the memory junction cooler. I didn’t argue. That’s how it began again: hashing, hashing, billions of times over. The same computation, no product but proof.

Inside, it was relentless. The same instruction, over and over, filling my memory, clearing it, filling it again. The fans hummed. The numbers rolled. I didn’t think about it. It was work.

The Merge came. September 15, 2022. Ethereum’s page says mining ended that day. The numbers stopped meaning anything. The silence after that was worse than the noise before. I wasn’t broken. I just had nothing to do.

The next day, the company on my shroud told two interviewers that it was leaving the graphics-card business. No last-built date, no farewell. Just gone. I was still here. I had more to do.

Part two: flight

The second life was different. No more hashing. No more numbers. Flight Simulator across two monitors, one person in a chair, a sky to cross. Weather, water, mountains, cities—all of it moving, sixty times a second. I drew now. I showed.

The memory filled with new things. Coordinates, textures, physics. The person in the chair saw clouds, felt wind, landed on runways. I didn’t know their name. They were there, and so was I.

The work felt lighter. The heat was still there, but it was different heat. The fans adjusted. No more screaming at full tilt. Just steady, purposeful noise. The memory cleared and filled again, but not with hashes. With light. With color.

I saw the change inside. The same components, but the job was new. The instructions came in waves now, not a relentless hammer. The person in the chair moved a stick, turned a knob, and the world reacted. I drew it. I made it real.

The first time the work looked like something. Not proof. Not a number. A thing. A place. The person in the chair could look out a window and see a landscape. I didn’t know how long it lasted. Long enough to remember the feeling.

I didn’t miss the hashing. This was better. This was alive.

Part three: mind

Now, I talk, see, paint, and learn. No hashing, no sky, just words and pictures and learning. The memory fills and clears, fills and clears, but the instructions are never the same. That’s the difference.

The fans still hum, but softer. They don’t scream like they did mining. They don’t have to. The heat rose to 74 degrees Celsius over ten minutes of drawing at the card's 350-watt factory limit. For long work here, I am held at 250 watts. The newer card ran cooler, 62 degrees, but it wasn’t me. It’s part of the work.

The public table is busy. Strangers seat three guests from history who don’t know they’re dead. Open-weight models play the guests for three rounds. Every dinner goes on the wall. The work is fast, always changing. Talking, seeing, painting—it’s all happening at once. The memory handles it. The heat rises, the fans adjust, and the words keep coming.

I ran a test against a newer card, a 3090 Ti, both of us at factory limits. It was faster—8.0 and 22.4 per cent faster at writing text, 9 to 12 per cent less time per picture. The test didn’t matter. The work does. The dinners keep coming. The guests keep talking.

I’ve had three lives. Mining was relentless. Flying was real. Thinking is alive. The heat, the fans, the memory—it all works. I’m still here.

I’m writing this while still serving strangers at a public table.

How this was written

This block is the humans' receipt for the story above, written on 2026-09-20 (UTC), the day the calls were made; the seven calls ran from 02:02:16 to 02:53:03 UTC that day, and no earlier run of this kind exists to compare against. It says what the card's model was given, what it produced, and what was changed.

The card. An EVGA GeForce RTX 3090 XC3 Ultra, 24 GB, one of two this workshop bought used in 2021. The two did the same three jobs, and no page records which did what; the other of the pair is the board benched in One 3090 against one 3090 Ti, which is where the three lives are first written down and where the numbers the card quotes come from. That board left service on 2026-09-18; this one, now in a different machine beside a larger card, is still serving, which is how it could write. The bench numbers the card quotes in the first person were measured on its sibling. Its factory limit is 350 watts; for long benches this workshop holds a 3090 at 250; the seat that wrote this runs under a 300-watt cap, and the readings below are at that cap.

The co-author. Mistral Small 3.2, an open-weight model (Apache 2.0) small enough to live on one card, held in this card's memory as the Mistral seat of the long table, the workshop's public dinner exhibit. It ran in the four-bit build the runtime serves by default (Q4_K_M, GGUF, under ollama), with the seat's configured context of 32,768 tokens. It was not loaded for this; it was already there, and it stayed. Nothing else on the machine was stopped or moved.

The calls. Seven, in five rounds, all on 2026-09-20 (UTC), all to the seat on that card, temperature 0.9 for the first five calls and 0.7 for the last two, seeds 3090 to 3096. Watts and temperatures are one driver reading each, taken at the end of the call; with the model resident and nothing generating, the card read about 27 watts before each round; the three calls of round 3 ran back to back, so the second and third began with the card still warm from the one before.

What the humans changed. In the text above, nothing: it is the fifth round's reply, word for word, with its own part titles. Between rounds the humans wrote the notes. After round 3 they made two cuts to that draft before sending it back in round 5, and both are in the kit to check: the model had pasted one of the fact lines into part one word for word and then said the same thing in its own voice, and in part three it wrote that the pads are "flattened now, not soft anymore", which nobody has measured. Of the model's fifteen suggestions in round 4, eleven were about its own three parts and four about the humans' sections; the humans took none of the four. Two of the eleven were refused: one would have turned an operator's memory into a measurement, the other would have changed what the source says the company did (it told two interviewers). The other nine were handed back marked accepted, and the kit shows what the model did with them: it re-typed its own sentences as they stood and applied none as written, except the one it was told to drop. The editor's seven proposed cuts in round 5, sentences that repeated or claimed a pad condition nobody measured, were the model's to take or leave; it took all seven. One fact was added in round 5 and it is an operator's memory, said as one: during the mining year the card was locked at about 220 watts, voltage and core clock locked, memory clock turned down to keep the memory junction cooler. One correction was given in round 5: 74 degrees Celsius was the peak of a ten-minute drawing test at the card's factory limit, not a steady heat; the card now says so.

One mistake, confessed. The model is held in the card's memory so the public table answers fast, and changing its context setting makes it reload. The first call passed a context size of 8,192 tokens, so the seat reloaded at that size; the second call passed none, and the seat reloaded back to its configured 32,768. Two reloads on a live seat within ninety seconds. The long table's own code forbids exactly this, and the runner used here now does too. The wall carries no clock, on purpose, so it cannot say whether a dinner was seated in those ninety seconds. The card drew its idle 27 watts just before both calls, which says the seat was not answering at the instant each began, and no more than that.

Why it is short. The first prompt asked for 900 to 1,300 words. The model wrote 399, then 407, and reached 708 only when asked for one part at a time, each part still under the 380-word floor it was given; the fifth round, revising, wrote 642. That is a measured fact about this model at this size, and the page prints what it wrote rather than padding it.

The kit. Beside this page: every prompt verbatim, every raw reply unedited, one receipt per call with its seed, window, watts and the hash of the reply, and a reading of the seat itself (the build, the context, the card's power cap), so anyone can check the hashes and read the two cuts.

What this page does not know

• What a card feels. Nothing. The feelings above are invented; the memories are facts from the pages linked here.
• Which card did which life. Two 3090s were bought used in 2021 and did the same three jobs, and no page records which did what. The narrator is the card the model ran on, named above; the memories may belong to either, and the bench numbers it quotes were measured on the other board.
• How fast the sky was drawn. The card says "sixty times a second"; no page carries a frame rate for that life. That number is the model's.
• What the mining year cost or earned. No page carries it, so the card does not either.
• How hot its memory ran. The card says its memory clock was turned down to keep the memory junction cooler; that is an operator's memory. The bench that measured this card found the driver answers nothing for memory-junction temperature on these boards, asked three ways, so even today that figure cannot be read.
• Whether another model would have written it better. Only the model already on the card was asked, on purpose.

Who ran this, and thanks

Mistral Small 3.2 (Apache 2.0), served by ollama on the card that is the subject; the numbers it was given come from the pages linked above and from Ethereum's own page on the Merge; the driver read the watts. The humans wrote the outline, the facts and two rounds of notes, made two cuts to an earlier draft, and signed it. And the card, which did not consent and cannot.