The budget I would defend in a room

What I expected

On Chapter 1 I wrote 8,000 MB as the budget. I had not yet measured the 1,498 MB floor. I expected the brain to be the large slice and all other slices to be small. I expected the last page to show a pie with a comfortable wedge left.

Thirty days ago, I also expected that the hard part would be to make models fast. The hard part was to make the numbers mine: measured, with the power mode recorded and the source written next to each number.

The theory

The budget as arithmetic, and the one line of physics under all of it

floor  1,498 / 7,620 = 19.7%
SLM    2,639 / 7,620 = 34.6%
STT + eyes  1,032 / 7,620 = 13.5%
left   2,442 / 7,620 = 32.0%
= 8,192 - 7,620 = 572 MiB reserved before Linux counts it

ridge = 1.25 TFLOP/s (fp32 spec) / 68 GB/s = 18 flop/byte
Q4 decode: 2 flop per 0.56 byte = 3.6
under the ridge: a token is a bandwidth read
1.2 B params x 0.56 byte = 0.67 GB
0.67 GB / 28 GB/s = 24 ms = 42 tok/s
= Day 4 measured 34.36: the gap, my guess, is launches

Unit check on the KV term: 144 KB per token (Chapter 3) x 4,096 tokens = 576 MiB, which is 604 MB. This term is a part of the 2,639 MiB SLM slice. The rest of that slice is weights and a CUDA context. Unit check on the pie: tegrastats prints MB but counts MiB, and the 8 GB chip is 8,192 MiB, so 572 MiB is reserved before Linux counts it. Each tegrastats and MemAvailable figure in Days 21 to 30 is MiB and is labeled MiB from this correction on. Chapter 1's first note said 380 MB, which was 8,000 - 7,620 across two units.

The setup

python3 build_day30.py
# reads every dayNN/measured.json; the pie is Day 25's residents

The measurement

Five bars, every one a tegrastats delta

The page has one pie of measured slices. It has three holes, and each hole names its instrument. It has one rule.

Where it broke

The wedge is not comfortable, and Chapter 27 showed it is not room for a VLA. SmolVLA is 1,602 to 1,744 MiB as a process. Beside the brain, the load took MemAvailable to 326 MiB and the guard stopped it. The budget does not close with all four resident and the desktop up. It closes as a choice: the SLM with a VLA, 6,853 of 7,620 MiB with 767 left, or the SLM with ears and eyes, 5,191 with 2,429 left in the stacked run (Chapter 25's own read of the same three was 5,178 and 2,442). In both cases Chapter 25's voice wants STT and TTS to take turns. Two reruns on 2026-10-04 stopped at 422 and 363.

The three holes are not small either. The loop that owns torque exists only in sim. Each camera number came from a synthetic source. The power mode, the clocks and the headless boot are each one command away. The walk-day date depends on two of the three. More inference work does not move them.

The first wrong number was 8,000. I wrote it before I measured anything. After that, I measured each slice before I put it in the budget.

Thirty days of runs measured inference and no walking. This pie is the brain half of the robot. The other half has no page yet.

Summary

• Three holes stay unmeasured after thirty days, each with its instrument named:
• the high-rate gait loop on real hardware: it runs at 1 kHz in sim (Chapter 29), not yet on real motors.
• the camera: each camera number in this series comes from a synthetic source (Days 6, 16, 20). The C920 changes the memory type and the p99
• the power mode: the harness recorded 15W and did not set it. No logged day shows MAXN or jetson_clocks. The headless boot, the systemd install and Nsight counters are each one command from a measurement.

Sources

These sources are not mine. Measured numbers came off my own board; specs, estimates and paper figures are labeled as such.

• The previous 29 pages
• Orin Nano 8GB Super datasheet
• Orin Nano 2 announcement as future hardware, not today's
• day27/rerun-2026-10-04/summary.json: the Day 25/27 stacked pile twice more on 2026-10-04