Hummingbird: World’s First Optical Network-on-Chip Accelerator for AI Workloads adaptable to Tensorflow
Lightelligence, the leading technology company, has unveiled Hummingbird™, an extraordinary Optical Network-on-Chip (oNOC) processor specifically designed for domain-specific artificial intelligence (AI) workloads. By leveraging advanced vertically stacked packaging technologies, Hummingbird integrates both a photonic chip and an electronic chip into a single package, marking a significant milestone in the quest to overcome the “memory wall” challenge faced by chip architects working on novel System-in-Packages with NoC interconnects.
Image Source – lightelligence.ai
Hummingbird serves as a cutting-edge communications network for data centers and high-performance applications, making it the flagship product in Lightelligence’s oNOC platform. By harnessing the power of silicon photonics, Hummingbird enhances computing performance by enabling innovative interconnect topologies. Utilizing waveguides that transmit signals at the speed of light, Hummingbird employs an all-to-all data broadcast network to efficiently communicate with each core on a 64-core domain-specific AI processor chip. This groundbreaking approach provides Hummingbird with significant advantages in terms of reduced latency and power consumption compared to conventional digital interconnect solutions.
One of the key differentiators of Hummingbird is its ability to achieve density scaling beyond the limitations of digital networks. Through the oNOC technology, Hummingbird enables interconnect topologies that were previously unrealizable, resulting in higher computing power utilization and more efficient communication. This breakthrough innovation simplifies the mapping of workloads to hardware, empowering users with greater freedom to select the optimal topology for their specific computing tasks. Hummingbird’s implementation as a low-latency optical all-to-all broadcast network spanning 64 cores, equipped with 64 transmitters and 512 receivers, provides a versatile framework for implementing a wide range of dense optical network topologies.
To ensure seamless integration, Hummingbird’s electronic and photonic integrated circuits (ICs) are co-packaged and integrated into a PCIe form factor, making it compatible with industry-standard servers. Moreover, when coupled with the Lightelligence Software Development Kit (SDK), Hummingbird empowers users to optimize machine learning and AI workloads, fully capitalizing on the capabilities of the oNOC technology.
Image Source – lightelligence.ai
Speaking about the significance of Hummingbird, a spokesperson from Lightelligence stated, “Hummingbird represents the world’s first compute product utilizing optical network-on-chip technology. It’s time to move beyond Moore’s law, which is reaching its physical limits, and focus on scaling the silicon area. Hummingbird’s exceptional performance unveils the next generation of computing architecture.”
Hummingbird’s adaptability to existing TensorFlow models further solidifies its position as a cutting-edge heterogeneous computing platform. Supporting nearly all TensorFlow operators, Hummingbird accelerates targeted graphs and dedicated operators, providing a powerful solution for AI and machine learning applications.
Image Source – lightelligence.ai
The 3D packaging architecture employed by Hummingbird showcases Lightelligence’s commitment to pushing technological boundaries. By seamlessly integrating a photonic IC and an electronic IC into a single package using advanced vertically stacked packaging technologies, Hummingbird exemplifies Lightelligence’s dedication to developing innovative solutions that shape the future of computing.
With the introduction of Hummingbird, Lightelligence establishes itself as a frontrunner in optical network-on-chip technology. This breakthrough product not only addresses the increasing demands of data centers and high-performance computing applications but also sets the stage for remarkable advancements in the field of AI computing.