A 0.66-Picojoule Magnetic Synapse Works at Room Temperature. A 2022 GPU Already Matches It.

Zero point six six picojoules: the projected energy per operation of a new artificial synapse reported in Advanced Materials by Jixiang Huang and colleagues at the University of Edinburgh, arriving wrapped in the usual "pave the way for energy-efficient AI" framing. Before deciding what it means, run the number against everything else on the ladder. A human brain manages about 20 femtojoules per synaptic operation. An NVIDIA H100 needs roughly 0.71 picojoules for one FP16 operation at the wall plug. Intel's Loihi burns about 20 picojoules per synaptic event. On raw per-operation energy, the new device does not beat the best silicon we already own. It ties it. Full stop. That tie is the most interesting thing about this paper, because it forces the question of what "energy-efficient AI" is actually supposed to mean.

Herding the whole herd

Skyrmions are nanoscale, vortex-like arrangements of magnetic moments that behave like stable particles. Physicists have been trying to turn them into artificial synapses for years, because they are tiny, nonvolatile, and move under small electrical currents. Until now the standard approach was always the same: create or destroy individual skyrmions to change the synaptic weight, the way a neuron strengthens or weakens a connection. That approach has a flaw the field has never solved: nucleating and annihilating individual skyrmions is stochastic. You get a different answer every time, and machine learning hates randomness in its weights.

Huang's team did something structurally different: instead of manipulating individual skyrmions, they drove a collective transformation of the entire magnetic texture, morphing a lattice of skyrmions into stripe-like magnetic domains inside a two-dimensional van der Waals ferromagnet called Fe3GaTe2, with thousands of magnetic moments moving together, deterministically. As the readout, the material's anomalous Hall resistance changes linearly and reproducibly with the duration of the applied pulse, giving multiple stable weight states and the multiply-accumulate behavior that neural networks are built on.

"Rather than manipulating magnetic skyrmions individually, we exploit their collective behavior," said lead author Elton Santos of Edinburgh's School of Physics and Astronomy. "This gives us a much more deterministic and reproducible way of controlling information while retaining the advantages of these remarkably small topological magnetic structures." Backpropagation, the algorithm that trains nearly every neural network, requires weight updates that are linear and symmetric, so linearity matters enormously here: this is the first skyrmion device to offer both without fighting physics. And the effect works at room temperature, which sounds mundane until you remember that most quantum and magnetic phenomena this exotic only survive near absolute zero.

The energy ladder, with the arithmetic shown

Nobody published this comparison, so here it is. Each rung is energy per operation, lower is better, and every figure below carries its inputs.

Energy per operation across devices, lower is better; derivations are in the text.
Device	Energy per operation	How the number is derived
Human brain	~0.02 pJ	20 W / 10¹⁵ synaptic ops per second
Analog Dynap-se	~0.1 pJ	Measured, per published neuromorphic review
NVIDIA H100, FP8	~0.35 pJ	700 W / 1.979 x 10¹⁵ ops/s (dense, no sparsity)
Skyrmion synapse (projected)	~0.66 pJ	Huang et al., scaled device projection
NVIDIA H100, FP16	~0.71 pJ	700 W / 9.895 x 10¹⁴ ops/s (dense, no sparsity)
NVIDIA H100, TF32	~1.42 pJ	700 W / 4.945 x 10¹⁴ ops/s (dense, no sparsity)
IBM TrueNorth	~22 pJ	46 x 10⁹ synaptic ops/s per watt, inverted
Intel Loihi	~20 pJ	Measured, per published neuromorphic review
SpiNNaker	~110,000 pJ	0.11 microjoules per synaptic event, measured

Read that table twice, because the projected skyrmion device sits 33 times above the human brain, about 7 times above an analog neuromorphic chip from a decade ago, essentially dead even with an H100 doing FP16 math, and roughly 30 times below Intel Loihi and IBM TrueNorth, the famous digital neuromorphic chips. A 2020 simulation of an earlier skyrmion racetrack synapse, by some of the same researchers, estimated about 1 pJ per synaptic event, so the field has improved roughly 1.5x in six years while adding room-temperature operation and determinism, which is respectable, but it is not a revolution.

Now the brain-scale thought experiment: give this device a brain's workload, 10¹⁵ synaptic operations per second, and multiply, 10¹⁵ x 0.66 x 10^-12 joules equals 660,000 watts, so a brain-scale skyrmion computer would draw 660 kilowatts, thirty-three thousand times the brain's 20 watts. Most of that gap is not device physics but firing rates, since biological neurons tick along at a few hertz while engineered devices can cycle millions of times faster, which means a brain-scale machine could in principle do brain-scale work much faster than a brain; still, thirty-three thousand is a sobering multiple.

Why it still matters

If the device ties a GPU on per-op energy, the honest case for it has to live somewhere else, and it does, in three places specifically.

First, room temperature changes the deployment physics, because cryogenic neuromorphic devices, like the silicon-carbide devices that mimic neurons at 10 millikelvin, need dilution refrigerators that cost more than the chip and drink kilowatts to stay cold, while a room-temperature magnetic synapse plugs into normal electronics. Second, nonvolatility means the weight persists with zero power, while a GPU forgets everything the instant you cut power and pays a data-movement tax every time it fetches a weight from memory, whereas a magnetic synapse stores the weight in the texture itself and computes where the data lives. Third, determinism is the unlock for training, not just inference, because stochastic skyrmion devices were always going to be inference-only curiosities, while a linear, reproducible device can in principle be trained in place, which is where most of AI's energy budget actually goes.

To test trainability the researchers did what the field always does: they fed the device's measured characteristics into a hardware-informed quantized neural network and ran MNIST digit recognition, hitting about 96.1 percent accuracy. Fine, and nobody disputes the device learns toy patterns.

What this does not prove

Remember that the 0.66 picojoule figure is a projection for future scaled-down devices, not a measurement on the lab device that was actually built, and every comparison in the ladder above inherits that projection, so treat the GPU parity as the device's best plausible future rather than its measured present. And the 96.1 percent MNIST result is a simulation informed by device data, not a fabricated crossbar chip running inference; software baselines for small MNIST networks sit around 98 to 99 percent, so there is a small accuracy tax for the hardware's imperfections. Comparing a synaptic operation to a GPU FLOP is deliberately provocative and only approximately fair: the GPU figures are theoretical peak throughput at the wall plug, a number real workloads never reach, while the brain's 10¹⁴ to 10¹⁵ operations-per-second estimate spans an order of magnitude depending on whose neuroscience you trust. Fe3GaTe2 is a van der Waals crystal grown in a research lab, not a material any CMOS foundry runs, and the public reporting includes no endurance or retention numbers for the collective transformation. We do not know how many times you can morph the texture before it degrades.

The strongest case against

Here is the steelman, stated at full strength. If a projected, best-case skyrmion synapse merely ties a 2022 GPU on per-operation energy, then "energy-efficient AI" is doing misleading work in the headline. Its advantages are architectural and materials-based: room temperature, determinism, nonvolatility, in-memory compute. Architectural advantages only materialize inside a complete system, and nobody has built one. Not one. A single Hall bar demonstrating a weight update is to a neuromorphic computer what a single transistor is to a CPU, which is to say a very long way away. Meanwhile the incumbent keeps improving on the metric that matters, energy per useful inference, and the memristor competitors this paper claims parity with, resistive RAM and phase-change memory, already sit at the same sub-picojoule table without requiring anyone to grow exotic 2D ferromagnets. Stripped of the press-release framing, the honest claim is narrower: this is the first room-temperature, deterministic magnetic synapse, and on raw energy it ties the best silicon we have rather than beating it. That is still a genuine first, even though it is not the first the headline implies.

What to watch

Nothing about this is actionable for a buyer today, and anyone selling you a skyrmion product roadmap is selling fiction. But the signals that would change that are specific and checkable. A fabricated crossbar array, not a single device, running inference on a real workload. Endurance data showing the collective transformation survives millions of cycles. Integration of Fe3GaTe2, or any 2D ferromagnet, with standard CMOS back-end-of-line processing at temperatures the rest of the chip can tolerate. For chip architects, the portable lesson is the design principle: collective, deterministic bulk transformations beat stochastic single-object manipulation for linear weight updates, and that principle generalizes beyond skyrmions. Edinburgh's own group separately reported cutting magnetic memory switching energy to 0.94 nanojoules with optimized field pulses, the same philosophy applied to a different problem. For investors, fund this as a room-temperature nonvolatile memory story, not an energy story; that energy story already belongs to the GPU. Already does.

The Bottom Line

Edinburgh's skyrmion synapse is a real materials advance, room temperature, deterministic, linear, projected at 0.66 picojoules, and it deserves to be covered as exactly that. But the energy ladder does not lie: at its projected best it ties a 2022 H100, sits 33x above a human brain, and its advantages only count inside a computer nobody has built. Here is the uncomfortable arithmetic: the device that beats your GPU on energy per operation already exists, and it is your GPU. The skyrmion's job is to become a different kind of computer, not a cheaper copy of this one.