Microsoft’s Light-Based Computer Could Boost AI Efficiency 100x
▼ Summary
– A new analog optical computer (AOC) uses light and voltage intensities instead of digital switches to perform calculations, offering a new computing paradigm.
– The AOC is estimated to be about 100 times more energy-efficient than traditional digital computers, particularly for AI and optimization tasks.
– It operates by repeatedly computing problems in a feedback loop until reaching a steady-state solution, avoiding analog-to-digital conversion to save energy.
– Researchers developed a digital twin of the AOC to handle larger and more complex problems, such as accurately reconstructing brain scan images with reduced data.
– The prototype successfully performed machine learning tasks, solved financial optimization problems with high success rates, and shows potential for scaling to handle millions of variables.
A groundbreaking computer that harnesses light rather than electricity for its operations promises to dramatically enhance the energy efficiency of artificial intelligence systems. Researchers at Microsoft have developed a prototype analog optical computer (AOC) capable of performing specialized AI tasks and solving complex optimization problems with far less power consumption than conventional digital machines.
This innovative system uses micro-LEDs and camera sensors to process information through light and varying voltage intensities, operating within a feedback loop that refines its results until reaching a stable solution. By avoiding the conversion of analog signals to digital format during computation, the AOC achieves an estimated hundredfold improvement in energy efficiency, a leap that could redefine hardware performance in AI applications.
Unlike general-purpose computers, the AOC functions as a dedicated “steady-state finder,” tailored for specific challenges in machine learning and optimization. A digital twin, a software model replicating the physical device, allows researchers to simulate larger and more intricate problems than the current hardware can manage directly.
In practical tests, the physical AOC matched the performance of digital computers in basic image classification tasks. Its digital counterpart successfully reconstructed a detailed brain scan using only 62.5% of the original data, suggesting potential applications in reducing MRI scan durations. Additionally, the system outperformed existing quantum computers in solving multi-party financial risk minimization problems, demonstrating its versatility and precision.
Although still in the prototype phase, the technology holds immense promise. Future iterations with expanded micro-LED arrays could handle millions or even billions of variables simultaneously, opening new frontiers in high-speed, low-power computing for AI and scientific research.
(Source: Live Science)