The breakthrough of the "Light Chip" in the US: A hundredfold Increase in AI Efficiency, Is it a disruptive breakthrough or excessive hype?
The power consumption for training a large AI model is enough to power thousands of households for a whole year. However, a tiny optical chip from the University of Florida claims to be able to reduce this energy consumption to just one percent of the original. Behind the 98% recognition accuracy rate lies a crazy game between computing efficiency and the energy crisis.
Release time:
2025-09-23
The power consumption for training a large AI model is enough to power thousands of households for a whole year. However, a tiny optical chip from the University of Florida claims to be able to reduce this energy consumption to just one percent of the original. Behind the 98% recognition accuracy rate lies a crazy game between computing efficiency and the energy crisis.
In September 2025, the University of Florida in the United States, in collaboration with multiple institutions, published an optic-based chip in the journal Advanced Photonics. This chip can complete AI tasks using lasers and microlenses, consuming only 1% of the energy of traditional chips while increasing efficiency by 10 times or even a hundred times. The team directly integrated optical components onto silicon chips, using optical signals to replace electronic operations, allowing convolution calculations (the core energy-consuming link in machine learning) to be completed directly during the light propagation process.
In the test, the accuracy rate of this chip in recognizing handwritten numbers reached 98%, which was on par with the performance of traditional electronic chips. More importantly, it adopts multi-wavelength laser synchronous processing technology - different colors of light represent different data streams, just like "VIP multi-lanes" running in parallel without any interference, significantly enhancing data throughput.
The speed is close to the physical limit, with no heat loss during propagation, and multiple signals do not interfere with each other. The three-dimensional optoelectronic chip developed by Columbia University and Cornell University in 2025 has a bandwidth density of 5.3 terabytes per second per square millimeter, setting a record for data transmission energy efficiency. The latency of photonic chips can be reduced to the nanosecond level, while that of electronic chips requires the microsecond level. The computing speed can reach 3200 TFlops, far exceeding the 7.8 TFlops of traditional chips.
Latest Blog