- Neurophos develops photonic AI chips replacing electrons with light for efficiency
- Pilot with Terakraft aims to demonstrate sustainable ultra-efficient optical computing
- Single chip promises 100 GPUs performance using 1% energy consumption
An American startup is preparing to test an experimental processor which could deliver performance at a scale never before achieved on a single chip.
Neurophos, which was spun out of Duke University with backing from incubator Metacept, is partnering with Norwegian data center operator Terakraft to run a pilot of its optical AI inference platform in 2027.
Combining photonics with metamaterials to shrink optical modulators by a factor of 10,000x, the company’s technology uses optical systolic arrays that replace electrons with light, removing latency bottlenecks while running at clock speeds above 100GHz.
Overcoming energy walls
By combining this with compute-in-memory architectures, Neurophos believes its chips could overcome the energy walls that limit conventional GPUs and TPUs.
Neurophos, which raised $7.2 million in late 2023, claims its technology will allow a single chip to deliver the compute power of 100 GPUs while using only 1% of the energy.
The planned collaboration will see Neurophos’ optical processing units deployed at Terakraft’s green data center in Norway. Previously part of the Sauda I hydropower plant (decommissioned in 2008), the facility ranks as one of the most efficient in the world.
“By hosting Neurophos’ ultra-efficient optical chips in our green data center for select enterprise clients, we not only reduce our carbon footprint but also raise the bar for energy-efficient AI infrastructure,” said Giorgio Sbriglia, chairman of the board of Terakraft. “Our mission has always been to power the future responsibly, and this collaboration brings that vision to life.”
Patrick Bowen, Neurophos founder and CEO added, “Terakraft’s commitment to renewable energy and innovative technologies aligns perfectly with our mission to democratize high-performance AI. By deploying our 100x more efficient inference chips in Terakraft’s green data center, we’re proving that AI’s exponential growth can be achieved sustainably, together.”
If everything goes as planned, the pilot in Norway could mark an early step toward sustainable ultra-efficient AI hardware designed to handle future workloads at scale.
Neurophos claims that end-to-end simulations validate its technology’s performance, with a roadmap targeting exaflop-class computing on a single chip.
Via eeNews Europe
