Efficient IT begins before the code

01Year 1 · Physical constraints

The first year establishes physical and infrastructural literacy. Students distinguish energy from power; examine mechanical, chemical, electrical, thermal and nuclear forms; and follow the transformations through which useful work is obtained. They encounter thermodynamics not as a decorative chapter of physics, but as a constraint on every energy system.

The first law states that energy is conserved. It does not state that every conversion is equally useful or reversible. Heat passes from a hotter body to a colder one; engines reject heat; transport and conversion introduce losses; the second law explains why a technically possible energy balance does not imply a reversible process. “Efficiency” therefore has a numerator, a denominator and a physical limit before it becomes a software metric.

The year then moves through heat and mass transfer, thermal machines, waves, remote sensing, fossil and nuclear energy, electricity and electromechanical conversion. A workshop on the energy optimisation of a motorway toll facility brings the layers together: AI is introduced after the physical system, its measurements and its constraints have been described.

02Grid & control

Electricity adds a constraint that computing students recognise immediately: the system must be observed and controlled continuously. Production and demand have to remain balanced. Too little generation is a problem; uncontrolled overproduction is also a problem. Variable wind and solar production, maintenance of large plants, storage limits and changing demand all affect the dispatch decision.

Robert Plana’s recorded classes make this operational dimension explicit. Dispatchable generation can be adjusted; storage can move some energy through time; interconnected networks can share capacity; smart meters and sensors can reveal a local disturbance before it propagates. The grid is physical infrastructure, but it is also a distributed sensing, forecasting, optimisation and protection problem.

That connection matters later. A computing workload is not supplied by an abstract quantity called electricity. It arrives at a particular place and time, through a network with finite capacity, from a mix of assets with different response times, availability, emissions and maintenance constraints.

03Year 2 · Coupled systems

The second year changes scale. Climate is not presented as a sequence of exceptional weather events, but as the statistics and dynamics of a coupled system. Weather describes short-term atmospheric conditions. Climate concerns longer periods, distributions and extremes, and requires the atmosphere, oceans, ice, land and biosphere to be considered together.

David Medio’s introductory lecture follows energy through that system. Solar and terrestrial radiation interact with gases, aerosols and clouds. Atmosphere and ocean exchange heat, water and momentum. Ice modifies albedo and sea level; vegetation exchanges carbon and water; ocean circulation redistributes heat across the planet. A change in one component can alter several others, sometimes over very different timescales.

This is where the seminar’s connection to data and computer science becomes methodological. A trend is not an anecdote. A model is not the system itself. Measurements have spatial coverage, uncertainty and historical limitations. A credible conclusion connects observations, statistics and a physically plausible mechanism. Climate science therefore gives students a demanding example of reasoning from incomplete data inside a nonlinear system.

04Year 3 · Computing decisions

Only in the third year does computing itself become the principal object of analysis. By then students have encountered cloud infrastructure, data engineering and AI. They can examine data centres, storage, networks and algorithms without treating electricity consumption as an isolated number.

The course considers computing in two roles. It is a load: processors, memory, storage, data movement, cooling, resilience and hardware replacement all require resources. It is also an instrument: the same technologies support climate modelling, remote sensing, smart grids, demand forecasting, water management, mobility and industrial optimisation. One role does not cancel the other.

The engineering question is therefore not whether AI or cloud computing is “good” or “bad” for the environment. It is whether a defined service produces enough value to justify its measured impacts; whether another architecture can provide the same service with fewer resources; and whether an apparent improvement shifts the burden to another layer, location or time.

05Metrics

Metrics are useful only when their question is explicit. Energy per inference, carbon per transaction, server utilisation and power usage effectiveness describe different parts of a system. None is a universal score. A lower value can coexist with a higher total impact when demand grows, when work moves elsewhere, or when greater efficiency makes a service cheap enough to be used much more often.

This is the rebound problem in practical form. It does not make efficiency pointless. It means that engineers must report both intensity and scale, and distinguish a local optimisation from a system-level reduction. They must also state whether estimates concern operational electricity, lifecycle impacts, average grid emissions or marginal changes.

06Why three years

The seminar lasts three years because the questions become more precise as the students’ technical vocabulary grows. In the first year, conservation, conversion and infrastructure constrain the problem. In the second, coupled systems, observations and uncertainty complicate it. In the third, computing architectures can be examined inside those physical boundaries.

The progression was proposed by José Massol, DSTI’s President and co-founder. His education at Arts et Métiers and Supélec, followed by a career in complex systems in the French defence industry, informed a straightforward curricular question: can a computer engineer be fully trained while remaining unaware of the energy and physical infrastructure on which digital systems depend? The teaching and scientific content are developed by the professors responsible for each part; the original contribution was to insist that the question should remain present throughout the degree.

07Contributors

The credentials matter only because this subject crosses disciplines. The seminar is taught by people whose work places them inside the systems being discussed, not because the article needs a roll-call of titles.

08The useful result is a better question

The seminar does not end with an approved list of technologies. It should leave an engineer less willing to accept a claim whose physical assumptions are hidden.

Editorial note: this article synthesises the content and rationale of a seminar taught by several contributors across the BSc. It does not attribute the editorial wording to any individual lecturer, and it simplifies some scientific and engineering questions for a general technical readership.