University of Tokyo Researchers Accelerate Chips 1000x Without Overheating
A device based on magnetic switches achieves a thousandfold increase in data processing speed without generating additional heat. The technology promises a revolution in processors, though commercialization is still years away.
University of Tokyo Magnetic Switch: Why 40 Picoseconds Rewrite the Future of AI Chips
[The Gist]: What's Really Happening
A research team from the University of Tokyo, in collaboration with RIKEN, has created a non-volatile quantum switching element based on the antiferromagnetic material Mn₃Sn—a compound of manganese and tin. The device switches its magnetic state in 40 picoseconds, roughly 1000 times faster than the typical nanosecond operations in modern DRAM and AI accelerators. The results were published this week in the journal Science.
The real significance of the work is not the record speed per se. What matters is that the switching generates almost no heat—simulations showed a temperature rise of only 8 kelvins. For comparison, previous attempts to achieve picosecond switching caused temperature spikes of hundreds of kelvins, making them commercially unviable. The Tokyo group circumvented this problem through a spin-orbit torque (SOT) mechanism, where angular momentum is transferred directly to the magnetic structure rather than through material heating.
Practical implications: if the technology can be scaled to chip level, a Google-scale data center could consume energy equivalent to about 800 households instead of the current 80,000. Given hyperscalers' current spending on electricity and cooling—hundreds of millions of dollars annually per large cluster—the economic impact could be enormous.
Timeline and Context
January 2025. A paper in Nature lays the theoretical foundation for antiferromagnetic spin switches. The Tokyo group publishes preliminary results on manipulating magnetic states in Mn₃Sn.
2024–2025. The global semiconductor industry hits the thermal barrier. GPU clusters scale to hundreds of thousands of accelerators, making power consumption and cooling the main bottlenecks. DRAM requires constant charge refresh in capacitors—even idle systems consume energy for data regeneration.
May 2026. Publication in Science. The Tokyo group demonstrates a fully functional device: a layered Mn₃Sn/Ta structure on a silicon substrate, switching in 40 picoseconds, stable operation after more than 100 billion write cycles. Conventional chips at comparable speeds would overheat after just 10 million cycles.
Concurrently. The researchers demonstrated optical switching: 60-picosecond photocurrent pulses from a telecom laser and photodiode directly write information into the magnetic state. This opens the path to direct integration of optical data transmission channels with non-volatile memory—exactly the direction data center architectures with optical interconnects are moving.
Who Wins and Who Loses
Winners:
AI accelerator and data center manufacturers. NVIDIA, AMD, Intel, Google TPU—all face the challenge that further scaling of GPU clusters hits power and heat limits. A technology that reduces state-switching energy by a factor of 100 radically changes the unit economics of AI inference and training.
Japan's semiconductor ecosystem. The University of Tokyo and RIKEN cement Japan's status as a center of excellence in spintronics. Given government support programs for the semiconductor industry (Rapidus, TSMC Kumamoto subsidies), this gives Japanese companies a potential edge in the next generation of beyond-CMOS technologies.
IoT and edge AI device developers. Non-volatility—the magnetic state persists after power is cut—makes Mn₃Sn devices ideal candidates for autonomous sensors and edge AI devices where batteries must last for years.
Losers:
DRAM and NAND manufacturers. Mn₃Sn technology combines DRAM speed (but 1000x faster) with NAND non-volatility. If commercialization happens, the distinction between RAM and storage begins to blur—a fundamental threat to two pillars of the memory market worth over $200 billion combined.
Traditional approaches to ultrafast switching. All previous picosecond schemes based on thermal state destruction (heating by hundreds of kelvins) are now obsolete. The Tokyo group's work proves that a fundamentally different—non-thermal—mechanism is possible.
What the Media Isn't Saying
Most coverage focuses on flashy numbers—"1000x faster," "three months without charging a MacBook." But three systemic nuances remain underreported.
First. A 1000x increase in bit switching speed does not mean a 1000x faster computer. Real performance depends on memory, interconnects, chip architecture, and software. A fast switch is necessary but not sufficient. Professor Tomo Nakatsuji, the research lead, acknowledges this directly.
Second. The current implementation requires an external biasing magnetic field for deterministic switching. This is a serious practical limitation for a commercial chip—no one will place a permanent magnet next to every switching element. A solution to this engineering problem has not yet been found.
Third—and this is the most important underestimated aspect. The team demonstrated optical switching using a telecom laser. This means data can enter the device directly over fiber optics, without intermediate electronic conversion. For data center architecture, this eliminates an entire class of transceivers and their associated energy costs. The Mn₃Sn device is not just memory—it is a potential next-generation optoelectronic interface. This part of the work, rather than the record switching speed, may prove most transformative in the long run.
Forecast: Next 30 Days and 90 Days
30 days (through end of June 2026).
The Science publication will spark a surge of academic activity. Groups at MIT, Stanford, Cambridge, EPFL, and IMEC will attempt to reproduce the result on their own setups. Within a month, 3–5 preprints with verification or methodological variations will appear.
Corporate R&D labs will conduct due diligence. Intel Labs and IBM Research, with long-standing spintronics programs, will request Mn₃Sn structure samples and begin internal testing.
The venture community will react faster than academia. Startups in spintronics and antiferromagnetic devices—currently few—will receive inbound interest from deep tech funds like Khosla Ventures and Lux Capital. First seed rounds in "antiferromagnetic computing" may be announced by the end of June.
90 days (through end of August 2026).
The key question is scalability. The University of Tokyo device is a single switch the size of a laboratory sample. To become a chip, billions of such elements must be integrated on one die, solving issues of Mn₃Sn layer uniformity, deposition precision, and crosstalk suppression.
In parallel, the materials problem arises. Mn₃Sn is a compound with no standardized production process in the semiconductor industry. If Applied Materials or Tokyo Electron announce development of Mn₃Sn-compatible equipment, that will signal the technology's transition from lab to industry.
The research team itself acknowledges: a chip prototype is not expected before 2030, and commercial availability years after that. This is not a "tomorrow" technology but a planning horizon for long-term R&D programs. DARPA and European agencies tracking beyond-CMOS developments may initiate targeted funding programs with an estimated budget of $50–100 million over the next few years.
The most likely early commercialization path is not full processors but embedding Mn₃Sn switches into existing photonic integrated circuits as modulators and optical memory elements. This scenario could materialize within 5–7 years, well before a "fully spintronic CPU," and deliver first energy savings in data centers on the order of $200–400 million per year for a large hyperscaler.
— Editorial Team
No comments yet.