PYTHON · NUMPY

What the table would not say

A published table gives a device's inference latency as 102.3 ms and its energy as 2.398 kJ, and does not say what the energy is per. What can you conclude?

Recover an unstated denominator from physics, on a real published cross-vendor energy table, and then show what demanding a control rate costs in joules.

01Challenge

A 2026 preprint measures six vision-language-action models across seven accelerators with a physical power meter, and publishes a leaderboard of cost, energy and time. The measurements are the authors' own and the method is the right one. Its energy column is in kilojoules and the table never says per what. Decide whether the column can be used anyway.

Predict firstA published table gives a device's inference latency as 102.3 ms and its energy as 2.398 kJ, and does not say what the energy is per. What can you conclude?

02Model

Lesson one asked what a number is per. This one is what happens when the number is real, measured carefully, published by people doing the right thing, and still does not say. The table is genuine: six models, seven accelerators, energy taken with a physical power meter, and the inference hardware and the robot arm measured separately, which is the separation this whole institute argues for. Its cost and time columns are immediately usable and they already carry a surprise, because the fastest part, the cheapest part and the lowest-energy part are three different parts and none of them wins twice. The energy column is the problem. A kilojoule is not a property of a device, it is a property of a device doing a stated amount of work. So there are two moves available. Refuse the column, which lesson six would let you do. Or notice that a wrong denominator implies a power the hardware could not draw, and let physics hand the denominator back. You will make the second move, and then you will find the thing the table can tell you that no single column of it could: what it costs, in joules, to insist on a faster control loop.

03Guided practice

1. Read the three rankings first. They are computed for you in the given cell. Notice that the answer to which part is best depends entirely on which column you were asked about. 2. Two conversions, one division. A kJ is 1000 J and a ms is 0.001 s. The elapsed time is n_inferences * latency_ms, not the latency alone. Check yourself against one kJ over one second, which is 1000 W. 3. Let the band do the work. The bracket is deliberately wide, 5 W to 700 W, so a denominator it rejects is rejected by physics rather than by a tight guess about any one part. 4. Then read the control-rate table. The chunk size comes from the model, not from the hardware, and it is what converts a latency in milliseconds into a loop you can or cannot close.

04Feedback

PASS WHEN You recovered a denominator the authors did not print, using nothing but their own two columns and a bracket wide enough to be uncontroversial, and then priced a control rate in joules.

If your run fails, check:

Out by a factor of 1000 means one of the two conversions is missing. A kJ is 1000 J and a ms is 0.001 s.
The elapsed time is n_inferences * latency, not the latency on its own. Dividing by the latency alone gives the same answer for every N, and the sweep will be flat.
Define implied_power(energy_kj, latency_ms, n_inferences) returning watts.

05Retrieve & space

The authors measured carefully with a real power meter and still shipped an unusable column. What does that tell you about where reporting fails, as against where measurement fails?
You recovered the denominator to one decade, not to a number. Say precisely what you are entitled to claim with the energy column now, and what you are still not.
At 250 Hz the best energy available is 1.871x the table's minimum, and no hardware changed. Where did that factor come from?

06Mastery & project

You can take a published figure whose denominator is missing, decide whether it is recoverable rather than merely absent, recover it where physics allows, and state the bound you are left with instead of the number you wanted.

Find a published energy or efficiency figure in your own field and try to invert it the same way. Write down the denominator the author omitted, the physical bracket you used, and whether the bracket narrowed the answer to one decade, to several, or to none. A figure that survives every denominator is not a measurement of anything.

GO FURTHER · THE LIVE PLAYGROUND

The leaderboard ↗ · Zhou, Chen, Peng, Li, Li and Gu, 2026: six VLA models across seven accelerators, energy taken with a physical power meter, hardware and arm measured separately.
What the number is per ↗ · The first lesson of this course, and the defect this table has in the wild.
Bounding what you cannot measure ↗ · The same move applied to a term that was left out rather than to a denominator that was.

← Bounding what you cannot measure Back to course →

PYTHON · NUMPY

What the table would not say

One real published table, one missing denominator, and one question it can answer that none of its columns can.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you