AES Semigas


1 March 2023

Ayar demos first 4Tbps optical solution, paving way for next-gen AI and data-center designs

Silicon photonics-based chip-to-chip optical connectivity firm Ayar Labs of Santa Clara, CA, USA has announced public demonstration of what it says is the industry’s first 4 terabit-per-second (Tbps) bidirectional wavelength division multiplexing (WDM) optical solution at the Optical Fiber Communication Conference (OFC 2023) in San Diego (5–9 March). The firm adds that it is currently working with high-volume manufacturing and supply partners including GlobalFoundries, Lumentum, Macom, Sivers Photonics and others to deliver the optical interconnects needed for data-intensive applications. Also at OFC, partner Quantifi Photonics is launching a CW-WDM-compliant test platform that integrates Ayar’s SuperNova light source.

Ayar says that in-package optical I/O uniquely changes the power and performance trajectories of system design by enabling compute, memory and network silicon to communicate with a fraction of the power and dramatically improved performance, latency and reach versus existing electrical I/O solutions. Delivered in a compact, co-packaged CMOS chiplet, optical I/O becomes foundational to next-generation artificial intelligence (AI), disaggregated data centers, dense 6G telecoms systems, phased-array sensory systems and more.

“In-package optical I/O solutions have the potential to transform how semiconductor, AI, HPC and aerospace customers process their next-generation, data-intensive workloads,” comments Craig Thompson, VP of business development for Networking at NVIDIA. “NVIDIA’s accelerated computing platform is enabled by advanced technologies such as WDM optical interconnects to equip tomorrow’s innovators with the extreme performance they need.”

At OFC, Ayar is giving a first public demonstration showing its optical I/O solution moving data from one TeraPHY optical I/O chiplet to another at 2.048Tbps each direction powered by its SuperNova light source. SuperNova powers eight fiber links (using 64 highly accurate wavelengths operating at 32Gbps, for eight wavelengths and 256Gbps per individual fiber) running error free at lower than 10ns of latency and without needing forward error correction (FEC). This allows for a total bandwidth of 2.048Tbps each direction, or 4.096Tbps bidirectional. More importantly, the data transfer is using less than 5pJ/bit (10W), a high level of energy efficiency, providing the power density and performance per watt needed to achieve AI models with trillions of parameters, advanced HPC designs and more.

“Ayar Labs continues to showcase our technology leadership with this live silicon demonstration, an industry-first milestone on the path to overcoming the impending power and performance wall of electrical design, and unleashing the power of next generation compute,” says CEO Charlie Wuischpard. “As we bring together all the supply, manufacturing, test and compute pieces needed for high-volume deployment, we also show today that we continue to lead in pure technical achievement,” he adds.

According to the report ‘Co-packaged Optics for Datacenter’, revenue generated by optical I/O for high-performance computing (HPC) is increasing at a compound annual growth rate (CAGR) of 74% from about $5m in 2022 to $2.3bn in 2033, says Martin Vallo, senior analyst, Photonics, at Yole Intelligence (part of Yole Group). “Projections of rapidly growing training dataset sizes underline that data will become the main bottleneck for scaling machine learning (ML) models, resulting in a potential slowdown in artificial intelligence (AI) progress,” he adds. “Using optical I/O in ML hardware can help overcome this bottleneck, and could be the main driver for the adoption of optical interconnects for next-gen HPC systems.”

See related items:

Live demo at ECOC of Ayar’s SuperNova multi-wavelength optical source leveraging Sivers’ CW-WDM-compliant DFB laser array

Tags: Optical communications silicon photonics