Photonic computing, a paradigm that harnesses light instead of electrons for information processing and computation, is taking significant strides towards becoming a reality. Two recently developed computer chips, leveraging the unique properties of photons, have successfully tackled complex computing tasks that were once considered beyond the reach of purely photonic systems. These advancements demonstrate the potential of photonic processors to offer faster analog processing with significantly lower energy consumption compared to traditional electronic counterparts.
The fundamental advantage of photonic computing lies in the nature of light itself. Unlike electrons, photons do not interfere with each other, enabling highly parallel processing of information. This inherent parallelism promises a significant speedup for certain computational tasks. However, challenges such as noise in analog photonic systems and the difficulty of integrating them seamlessly with existing computing infrastructure have hindered their widespread adoption for over a decade.
The research team addressed specific photonic computing optimization problems by creating a chip called PACE (Photonic Arithmetic Computing Engine). This innovative design combines photonic components with electronic circuits to mimic the Ising spin model, a physical system used to frame a wide range of real-world optimization problems like image matching and network partitioning. The PACE chip employs a feedback loop, performing matrix-vector multiplication optically and then electronically processing the result to update the variables in the photonic processor. This iterative process allows the system to converge towards an optimal solution. In a compelling demonstration, the PACE chip achieved speeds 500 times faster than a modern graphics processing unit (GPU) in reconstructing a noisy image and solving optimization problems, showcasing the impressive computational power achievable through optical parallelism.
Meanwhile, another research team developed a photonic computing system capable of running standard artificial intelligence (AI) models like ResNet and BERT. Their approach utilizes photonic tensor cores, specialized units that perform matrix-vector multiplications using light instead of electricity. These photonic cores are controlled by accompanying electronic interface chips, demonstrating the feasibility of building a non-transistor-based computer capable of handling state-of-the-art AI workloads. Recognizing that data movement between processors is a major bottleneck in current AI systems, the team designed their digital interface chip to ensure fast and efficient communication with the photonic tensor cores.
To combat the inherent noise in analog photonic computing systems, which can significantly impact the accuracy of complex calculations, particularly in AI, the team implemented active calibration to dynamically adjust the chip’s power in response to light intensity fluctuations. They also introduced a specialized number format called adaptive block floating point. This clever mathematical solution groups numbers to share a common exponent, simplifying the hardware and effectively reducing noise without sacrificing the precision required for accurate AI model predictions. The system has successfully performed image classification, natural language processing, and even reinforcement learning tasks, running standard AI models without requiring any retraining or adaptation to the photonic hardware. The chip demonstrated a 7- to 10-bit precision, achieving comparable speeds to traditional GPUs on standard AI tasks while boasting at least ten times greater energy efficiency.
These two significant demonstrations highlight that photonic computing is transitioning from the realm of research labs to tangible real-world relevance. By tackling vastly different computational challenges with tailored photonic architectures, these advancements showcase the increasing scalability, reliability, and practical applicability of photonic processors, marking an “incredible leap ahead” in the field.
Related Content: Bistable Nanoparticles Redefine Optical Switches