Physicists at the University of Edinburgh have found a way to flip a magnetic memory bit using several orders of magnitude less energy than the technologies computers rely on today, by changing not how hard the magnetic field pushes but precisely how it changes over time.
The work, led by Dr Elton Santos of the Institute for Condensed Matter Physics and Complex Systems, was published in Advanced Materials on September 6.
Shaping the pulse instead of raising the power
Writing data to magnetic memory means reversing the magnetization of a tiny region of material – turning a zero into a one. Conventional approaches apply a field or a current strong enough to force the flip, and most of the energy spent goes into overcoming the material’s resistance to changing state rather than into the change itself.
The Edinburgh team took a different route, applying optimal control theory – a mathematical framework for finding the most efficient way to steer a system from one state to another – to design the shape of an ultrafast magnetic pulse. Rather than a blunt push, the pulse follows a trajectory calculated to work with the material’s own dynamics.
“By carefully designing how a magnetic field changes in time, magnetization can be switched far more efficiently than with conventional approaches,” Santos said.
Approaching the thermodynamic floor
The result is measured against the Landauer limit, the fundamental minimum energy that thermodynamics says is required to process a bit of information. It is an extraordinarily small number, and real memory technologies operate far above it – DRAM, the working memory in every computer, by an enormous margin, and the newer magnetic technologies STT-MRAM and SOT-MRAM by a smaller but still large one.
Approaching that floor is what makes the claim significant. The researchers report reductions of several orders of magnitude against all three of those technologies, and the framework was built to account for real experimental limitations rather than assuming ideal conditions – which is where theoretical efficiency gains usually evaporate.
Why memory energy matters now
Memory has become one of the dominant energy costs in computing, and the reason is AI. Training and running large models moves enormous volumes of data between memory and processors, and moving data costs substantially more energy than computing on it once it arrives. Data centre operators have spent the past several years discovering that their power budgets are shaped as much by memory traffic as by raw processing.
A reduction of the magnitude described here would not translate directly into a proportional cut in a data centre’s electricity bill – memory switching is one component among many, alongside cooling, interconnects and the processors themselves. But it addresses a cost that scales with every bit written, in a sector whose bit volumes are growing faster than any efficiency curve has kept up with.
What it is, and what it is not
This is a physics result, not a product. Several things stand between it and a memory module:
- Pulse generation. Producing precisely shaped ultrafast pulses in a laboratory is one problem; producing billions of them per second on a chip, cheaply, is another.
- Material scope. Results demonstrated in specific magnetic systems have to hold across the materials a manufacturer can actually deposit at scale.
- Integration. Memory technologies compete on density, endurance, retention and cost as much as on write energy, and a win in one column has to avoid losses in the others.
The researchers note the framework is not limited to magnetic fields. The same optimization applies to switching driven by electrical currents and by ultrafast laser pulses, which broadens the range of future storage technologies it could inform – and matters because current-driven switching is what commercial MRAM actually uses.
The longer arc
MRAM has been the perennial almost-technology of memory research for two decades: non-volatile like flash, fast like DRAM, and never quite cheap or dense enough to displace either. It has found real niches in embedded systems and industrial applications, and write energy has been one of the persistent obstacles to a broader role.
Work of this kind does not resolve that on its own. What it does is move a constraint that had been treated as material physics into the domain of control engineering – the flip costs what it costs because of how the field is applied, not only because of what the magnet is made of. That is a more tractable problem than the one the field started with.
