← Building the energy compute future · The Energy Lab
Two meanings, and the field needs both.
In thermodynamics, work is what a machine actually delivers for the energy it draws — the useful part, as distinct from the heat. In a career, work is what a person decides to spend their attention on. A computing paradigm that wants to exist in ten years needs both senses settled: a unit that states honestly what was accomplished per joule, and a set of open problems concrete enough that somebody can begin one this week without permission.
Work per joule, not operations per second.
Every era of computing has been organised around a number, and the number selects what gets built. Clock speed gave us deeper pipelines. Operations per second gave us wider matrix units. Both are measures of motion rather than of accomplishment, which is why a machine can improve on them while doing nothing more useful.
Cycles per second
Rewards frequency. Says nothing about whether the cycle did anything.
Operations per second
Rewards throughput. A machine can double it while fetching the same operands twice as often.
Operations per watt
Better, but still counts motion. Two chips can match on it while one solves the task and the other does not.
Work per joule
Names the task, the energy spent, and how the figure was obtained. The only one of the four that a claim can be checked against.
The difficulty with work per joule is that it cannot be reported as a single number, and that is a feature. It requires naming the task, so that two machines are compared on the same thing; the energy, measured or modelled and labelled as such; and the conditions, because a figure from a laboratory bench and one from a device in a field are different claims. A unit that resists being reduced to a headline is doing its job.
Open problems, stated so someone can start.
None of the following requires new silicon, an invitation, or a budget. Each is a piece of the field that is genuinely missing rather than merely unpublished, and each can be attempted by a person with a laptop and some patience. Where the Institute has a view on how to begin, it is stated; where nobody knows, that is stated too.
Measure a joules-per-operation table for one device class
Energy models for browser and edge workloads are currently estimated from architecture rather than measured. Take one device, sweep a known operation across occupancy and clock, regress the slope, and publish the table with its uncertainty. This single artifact converts a large amount of modelled work into something calibrated.
who: Anyone with a device, a power meter, and a fortnight.
Reproduce a published energy claim, and report what you find
The field's biggest weakness is that almost nothing is independently checked. Pick any claim from the map, attempt it, and publish the result whichever way it falls. A failed reproduction is more valuable than another announcement, and is currently much rarer.
who: Anyone who can run the workload. No permission needed.
Build a benchmark for learning under drift
The open capability is adaptation while running, and there is no accepted way to score it. It needs sealed environments that change on a schedule, a regret measure, a latency bound, and a mandatory comparison against the entrant's own frozen-weight ablation. Without that last rule, the benchmark measures the problem rather than the method.
who: Someone comfortable with reinforcement-learning evaluation.
Port one settling problem to a physical substrate and report the composed cost
Take a decision problem, run the settling part on hardware that settles and the rest on a GPU, and report one honest receipt across both. The interesting result is not the speed; it is whether the plumbing can be made to work and what the composition costs.
who: Someone with access to annealer or FPGA time.
Write the operators that make an interchange format cover this work
Model formats today assume dense layers and attention. A settling program needs operations that do not exist in them yet. The specification work is drafted; the implementation is not.
who: A compiler or runtime engineer.
Teach it. Build the explorable you wish you had had.
Most of this material is currently locked in papers and spreadsheets. A working simulation that lets somebody discover the memory wall for themselves does more for the field than another review, and takes an afternoon.
who: Anyone who can write a hundred lines of JavaScript.
These are open in the ordinary sense: nobody owns them, and doing one does not require joining anything. If a piece of this becomes your work, the Institute's interest is that the result exists and is checkable, not that it carries our name.
Fewer of one kind than people assume.
The stereotype is that a new computing paradigm needs device physicists. It needs some, but the shortages are elsewhere, and most of the missing work is closer to ordinary engineering than the subject's reputation suggests.
Measurement engineers
The single biggest gap. Almost every number in this field is modelled, and turning modelled figures into measured ones is unglamorous, tractable, and currently rate-limiting.
Compiler and runtime people
A substrate nobody can target is a substrate nobody uses. The history here is that portability layers decide which hardware survives, and they are written by people who like type systems, not lasers.
Benchmark designers
Choosing what to score decides what gets built. The field currently has no accepted way to measure its central claim, which is an opportunity disguised as an absence.
Teachers and explainers
Two hundred million simulation uses a year happen because somebody made physics explorable. The equivalent for this material does not exist yet.
Start with the course →See who is already working on itRoles at the Institute