PYTHON · NUMPY

Which block was that number measured over?

One chip, two published figures: 346 TOPS/W for the system and 8791 TOPS/W for its multiply-accumulate array. What does the gap tell you?

Turn two published efficiency figures for one part into the fraction of energy the headline one hides, and state the boundary a TOPS/W claim is really about.

01Challenge

A 2026 vision chip publishes two efficiency figures, both the authors' own, both honest, differing by more than twenty-five times. Work out which one describes the joules you would actually pay, and how much of the energy the other one cannot see.

Predict firstOne chip, two published figures: 346 TOPS/W for the system and 8791 TOPS/W for its multiply-accumulate array. What does the gap tell you?

02Model

This course has been about what a sense is worth, and the awkward fact underneath it is that the sensing term is usually missing from the figure people quote. Here is that happening on a chip built specifically to fix it. The part is a vision system-on-chip whose whole contribution is that the signal never becomes digital between the photodiode and the classification: the analogue-to-digital conversion is deleted rather than made smaller, and in a conventional pipeline that conversion and the pixel readout are most of the sensing budget. So this is the silicon that prices the sense. It publishes two efficiency numbers, both the authors' own and both accurate: one for the system and one for the multiply-accumulate array. They differ by more than twenty-five times, and because efficiency in operations per joule is a reciprocal of energy, that factor converts straight into a share of the joules. What you will find is that the array figure, the larger and more quotable one, is a measurement of about four percent of the energy, and that most of the ninety-six percent it cannot see is the sensing the chip exists to account for. The lesson is not that anyone is lying. It is that operations per watt is never a property of a part. It is a property of a boundary somebody drew, and a figure that does not name its boundary is not yet a claim.

03Guided practice

1. Reciprocals, not differences. Operations per joule means energy goes as ops / tops_w. Two figures for the same op count give two energies, and the ratio of the figures is the ratio of the energies inverted. 2. One line. The fraction outside the block is 1 - system_tops_w / block_tops_w. Check yourself on a block twice as efficient as the system, which must give exactly one half. 3. Sanity, both ends. A block that is the whole system must return zero. If your function can return a negative number, the two figures you were given are not nested and the comparison is meaningless. 4. Then read the sweep. It repeats the calculation for array figures from three times to a hundred times the system figure, so you can see which part of the conclusion depends on the exact number and which does not.

04Feedback

PASS WHEN You converted two efficiency figures into a share of the joules, found that the quotable one measures under four percent of the energy, and can now say what a TOPS/W number is actually a property of.

If your run fails, check:

Return a fraction, not a percentage. A block twice as efficient as the system must give 0.5, not 50.
The relationship is a reciprocal. If you subtracted the figures rather than dividing them, the units are operations per joule and the answer is not a fraction of anything.
Define energy_outside(system_tops_w, block_tops_w) returning one number.

05Retrieve & space

Both figures are the authors' own and both are accurate. Say precisely what makes the larger one unusable as a cost, without calling it wrong.
The sweep shows 90% hidden at a ten-fold gap and 99% at a hundred-fold. Which part of the lesson's conclusion survives not knowing the array figure at all?
This chip deletes the A/D conversion from the sensing path. Explain why quoting its array-only efficiency undoes the thing the chip was built to do.

06Mastery & project

You can read any operations-per-watt claim as a statement about a boundary rather than about a part, convert a gap between two such figures into the share of energy the narrower one omits, and say which of the two a buyer would actually pay.

Find an efficiency figure published for a part you might use, in TOPS/W, GFLOPS/W or joules per inference. Write down the boundary it was measured over, in words, and whether the sensor and the data movement to it are inside that boundary. If the source does not say, write that down instead: an unnamed boundary is the finding, and this course has taught you what it can cost.