Live · refreshes every 30 s · polled every minute

Coppelius training on the RTX 3090.

A chain of native training invocations on one rented card: invocation k trains on WikiText-103 context k (1,024 tokens, ten AdamW updates) resuming the checkpoint invocation k−1 wrote. Each row is one invocation the pod reported finished. This is an exercise of the mechanism on real text, not a claim of a trained model; see the device page for what the card is doing during an invocation.

Now · polled 2026-09-22T05:21:09Z

chain stopped by user after invocation 7 2026-09-22T05:05:26Z

8 of 30 invocations finished, 0 failed; mean wall 494.8 s. Card: NVIDIA GeForce RTX 3090, driver 580.126.09, 0 % utilization, 1 MiB, 32.14 W, 210 MHz, 37 °C at the poll. Pod disk 2.1 of 21 GB. Pod ofbo3f1ij3zgy3 (DONT-TOUCH-THIS-alpha-coppelius-3090-training).

GPU sample stream · last 720 × 500 ms of the running invocation

Utilization (solid) and power (dashed, of 350 W)

2 of 720 samples show utilization above zero. This chain ran with the clock-driven host (R8): 250 ms sleeps after every submission and two 120 s waits, so most samples were idle by design. The completion-driven host (R9) runs the same request in about two seconds.

Invocations

Every finished invocation

Predictions are the model's next-token guesses at rows 255, 511 and 1023 of its context after this invocation's updates; targets are the real next tokens. Hits so far: 0 of 24.

kExitWallCheckpoint SHA-256BytesPredictionsTargetsHits
70495.0 s5357405570548a8d692,465,664387 387 3871798 405 3640
60494.5 sb70cbf51deb0b78a692,465,6642101 387 3871798 290 640
50494.9 s3f420943ae430196692,465,6641064 1064 1064751 337 77860
40494.7 s92528ab41cacc3ee692,465,664670 1718 6351080 480 5670
30495.0 sb363cfbf6138aaa2692,465,664408 1337 1337265 369 60390
20494.8 sd2e55f21ddf0dc2f692,465,664670 1337 491080 2963 22560
10494.7 s1f53d0acc6a5adf3692,465,6642628 2779 2313784 858 6700
00494.7 seaffdd5ea375fd8a692,465,664265 158 4652646 1080 96270