HomeProductsProducts Details

AI chips with optical MAC and waveguide break memory-wall offering extreme HBM

Date: 02/07/2023
In a breakthrough Silicon-photonics technology, Lightelligence has unveiled its new optical waveguide integrated chip called Hummingbird, claiming as world’s first Optical Network-on-Chip (Onoc) processor designed for domain-specific artificial intelligence (AI) workloads for today's Hyperscale servers. Lightelligence, a company founded by X-MITs makes light signal processing a reality with Optical Multiply Accumulate (O-MAC) digital signal processor and optical waveguide based on-chip communication.

Hummingbird uses multiple stacked dies of optical waveguide fabricated photonic-chip and electronic-chip to increase data interface bandwidth between memory and processors and across processor elements, basicall clearing the bottleneck of memory and processor interface speed limits, what is now called as memory-wall enabling extreme HBM (High Bandwidth Memory). Lightelligence to demo its Hummingbird at Hot Chips August 27-29 at Stanford University.

“Photonics is the solution to the critical compute scaling problem, which has become pressing as the traditional solutions struggle to keep up with the exponential growth of compute power demand spurred by breakthroughs in the AI industry,” remarks Yichen Shen, CEO of Lightelligence. “Hummingbird demonstrates how the industry can address the scaling problem by incorporating photonic technologies into their next-generation product.”

“Lightelligence is breaking the memory wall with its proprietary photonics technology that could revolutionize the semiconductor industry,” adds Dylan Patel, Chief Analyst at SemiAnalysis.

Hummingbird follows Lightelligence' earlier 3D tech based Photonic Arithmetic Computing Engine (PACE) platform released in late-2021.

Hummingbird is the first from Lightelligence using its Onoc platform enabling silicon photonics. Semiconductor-made optically conducting waveguides to route light waves, are connected in a all-to-all data broadcast network topology to connect processing cores on a 64-core domain-specific AI processor chip enabling Hummingbird to break the memory-processor data limits offering high HBM severely reducing latency and power consumption replaceing copper/metal interconnect. Once metal interconnect is out on your high performance VLSI chip, it helps in further densing of processing elements.

Lightelligence explains "In Onoc, power and latency are virtually unaffected by distance, making the technology ideal for developing new and more robust topologies that do not rely on nearest neighbor communication. Onoc topologies like Hummingbird’s enable higher utilization of compute power even in a single electronic IC configuration due to more efficient communication. With Onoc, mapping workloads to hardware becomes easier and provides greater freedom to select the right topology for the computing task."

Lightelligence is one of the market's best silicon photonics chip having 64 transmitters and 512 receivers, where data-packets are broadcasted inside the chip across the processor cores via optical media.

To serve todays hyperscale AI server this new Hummingbird is integrated into a PCIe form factor ready for installation in industry-standard servers.

For quick development Lightelligence is offering Software Development Kit (SDK) to optimise machine learning and AI workloads. Onoc and Hummingbird IP can also be customized for other unique workloads and applications.

Lightelligence is planning to offer next versions of Hummingbird to support chiplet architectures to enable better scalability by using reticle-stitching.

Lightelligence said it is actively signing development partners to sample Hummingbird-based PCIe add-in cards along with Lightelligence's SDK in Q3 2023.

Lightelligence was founded in 2017 by Dr. Marin Soljacic, Dr. Yichen Shen, Dr. Huaiyu Meng, Spencer Powers, all from MIT.

Silicon can be effectively transformed into a optically conducting waveguides to route optical signal inside the chip. This author come across mutiple papers presented at events in this domain.
To learn more about this visit:
https://www.lightelligence.ai/index.php/tech.html