A Boston College-led research team has used a scanning nitrogen-vacancy quantum sensor to watch a new kind of transistor switch from inside the device. The experiment examined an atomically thin chromium sulfur bromide spin transistor that changes both its electrical transport and magnetic state, combining functions normally separated between logic transistors and nonvolatile memory.
The result matters most as a measurement advance, not as an immediate chip product. AI systems consume substantial energy moving model weights and intermediate data between processors and memory. A device that can compute and retain state in the same physical location could eventually reduce that traffic. The new work gives researchers a direct view of whether a candidate device is actually switching magnetically as intended, but it remains an early laboratory demonstration far from the scale, speed, yield and integration requirements of conventional computing hardware.[1]
By the numbers
- More than 1,000,000%: reported electrical on/off ratio for the chromium sulfur bromide spin transistor.
- 3,000%: reported magnetic on/off ratio.
- Atomically thin: the active material is chromium sulfur bromide, a layered magnetic semiconductor that can be exfoliated into ultrathin flakes.
- One device, two functions: the reported transistor is designed to pair switching logic with nonvolatile magnetic memory.
Why direct imaging changes the experiment
Most transistor measurements are indirect. Researchers apply voltage or current to a device and observe its electrical response at contacts. That is enough to establish that resistance changed, but it may not reveal why it changed, whether the response is spatially uniform, or whether the device’s magnetic state changed in concert with its charge transport.
The Boston College-led team addressed that gap with scanning nitrogen-vacancy, or NV, magnetometry. An NV center is an atomic-scale defect in diamond: a nitrogen atom adjacent to a missing carbon atom. Its electron-spin state is sensitive to local magnetic fields and can be read optically. By positioning a diamond sensor close to a sample and scanning it across the device, researchers can map magnetic stray fields with fine spatial resolution.
In this case, that capability allowed the researchers to correlate electrical switching with the magnetic behavior of the chromium sulfur bromide channel. The central significance is not simply that the device had a large electrical contrast. It is that the experiment connected a measurable circuit-level effect to the local magnetic configuration that could provide memory.
That distinction is important for spintronic devices. Electrical switching can arise from contact effects, charge trapping, heating, defects or material inhomogeneity. A simultaneous magnetic readout helps distinguish a claimed spin-based mechanism from a purely electrical artifact. The researchers reported that the transistor’s magnetic and electrical states can be switched together, with a more than million-percent electrical on/off ratio and a 3,000-percent magnetic on/off ratio.[1]
A transistor designed to store its own state
Conventional digital systems divide labor sharply. CMOS transistors perform logic operations, while SRAM, DRAM, flash and other memory technologies store data. Even when processor and memory are packaged closely, data must repeatedly travel across interconnects. That movement costs energy, adds latency and creates a system-level bottleneck often described as the memory wall.
The problem is especially acute for AI. Training and inference require repeated access to large arrays of weights, activations and key-value cache data. Improving model architectures and software efficiency can reduce those demands, but data movement is also a hardware and memory-architecture problem. If a device can hold a state after power is removed and also use that state in computation, it could support forms of in-memory or near-memory computing.
Chromium sulfur bromide, commonly abbreviated CrSBr, is attractive to researchers because it is a van der Waals material with intrinsic magnetic order in thin layers. Van der Waals materials consist of strongly bonded atomic sheets held together more weakly between layers, allowing extremely thin flakes to be isolated and combined with other materials. Their atomic-scale thickness also makes electrostatic control by a nearby gate especially effective.
In the reported transistor, electric control changes the device’s transport state while a corresponding magnetic state provides a nonvolatile degree of freedom. In practical terms, the ambition is a switch that does not merely represent a zero or one while power is applied, but can preserve useful state locally. That could reduce the need to write a result to a separate memory array and retrieve it later for the next operation.
The announcement’s two on/off ratios should be read as different measures of the same broader behavior, rather than interchangeable performance specifications. The electrical ratio describes the contrast in the device’s transport response; the magnetic ratio describes the measured magnetic contrast. Neither number alone establishes an application-ready energy-per-operation, switching speed, endurance or retention time. Those are separate metrics that will determine whether a device class can move beyond a physics demonstration.
What this could mean for AI hardware
The most credible long-term implication is not a drop-in replacement for CPUs, GPUs or standard memory. It is a possible building block for specialized architectures in which storage and computation are more tightly coupled. Such architectures could be useful for matrix operations, associative search, sensor processing or other workloads dominated by moving and repeatedly reusing data.
For AI hardware designers, the appeal of nonvolatile logic-memory elements is straightforward. A system could potentially preserve configuration or model state without continuous standby power, reduce transfers to distant memory, and make more decisions close to where data resides. The potential benefit is system-level: even a device that is efficient in isolation has limited value if packaging, converters, control circuitry and memory interfaces dominate total energy use.
The quantum microscope is also relevant beyond this particular material. Spintronics and two-dimensional electronics often involve nanoscale magnetic domains, interfaces and defects that conventional electrical probes average together. NV-based imaging gives device researchers a way to see where a state begins, whether it propagates uniformly, and how it changes under electrical control. That diagnostic role may prove more immediately useful than the single transistor itself because it can accelerate the screening and debugging of many candidate materials and device structures.
The work therefore fits a broader shift in advanced computing research: gains in AI efficiency will increasingly depend on devices, memory hierarchies, interconnects and packaging, not solely on making models smaller or software kernels faster. The transistor does not solve that problem by itself, but it demonstrates a mechanism researchers need to understand if they are to build hardware that attacks it.
Engineering hurdles remain substantial
There is a wide gap between observing switching in an atomically thin research device and manufacturing reliable arrays of billions of elements. CrSBr devices must be fabricated reproducibly, protected from environmental degradation where necessary, contacted with low resistance and integrated with control circuitry. Researchers must also show that their behavior remains stable across device-to-device variation and over long operating lifetimes.
Key unanswered engineering questions include operating temperature, switching voltage and energy, speed, endurance, retention, device footprint, error rates and compatibility with semiconductor manufacturing. A scanning NV microscope is an exceptionally capable laboratory probe, but it is not a production readout mechanism for a memory array. Any commercial implementation would need a scalable electrical method to write and read the relevant magnetic states.
There are also architectural trade-offs. Nonvolatile devices can reduce idle power and data movement, but AI accelerators require high throughput, dense integration and predictable analog or digital behavior. Added peripheral circuits can erase a device-level energy advantage. Large electrical on/off ratios are encouraging because they can support robust state discrimination, yet they do not by themselves demonstrate low-energy operation or high-volume manufacturability.
The announcement does not present independent industry validation or a commercialization timetable.[1] That is appropriate for work at this stage. The next meaningful evidence would be repeatable switching across larger device populations, measurements under application-relevant conditions, quantitative energy and speed benchmarks, and demonstrations of small integrated circuits rather than a single device.
What to watch next
The research establishes a useful feedback loop: electrical measurements show that the transistor switches, while nanoscale magnetometry shows how the magnetic portion of that switching behaves. That combination can help identify the material physics needed for future logic-memory devices and expose failure modes that would remain hidden in terminal measurements alone.
Near-term progress is likely to come from better control of interfaces, gates and material thickness, plus more complete measurements of how magnetic switching evolves across a device. Longer term, the decisive test will be whether similar behavior can be retained in dense, repeatable arrays that operate with practical voltages and temperatures and connect efficiently to conventional silicon logic.
For now, the work should be viewed as an enabling experimental result: it makes a promising category of spintronic transistor more observable and more falsifiable. That is a necessary step toward hardware that keeps data closer to computation, even if the route from quantum microscope images to energy-efficient AI systems is still long.
Editor’s Take
I see the quantum microscope as the most commercially relevant part of this announcement today. Device claims in emerging materials are often judged from terminal curves alone; being able to see magnetic behavior while the transistor is electrically switched gives engineers a much better way to separate a real operating mechanism from an impressive but misleading measurement artifact.
The million-percent electrical ratio is a strong laboratory signal, but it is not an AI-chip benchmark. I would watch for switching energy, speed, temperature range, cycling endurance and wafer-scale reproducibility before treating CrSBr as a contender for deployed accelerators. If those metrics advance, the logic-memory concept could matter because cutting data movement attacks a costly part of AI computing that model optimization cannot fully eliminate.
References
- EurekAlert! – https://www.eurekalert.org/news-releases/1136831
