Photowave, an Optical Networking (oNET) transceiver that leverages the latency and energy efficiency of photonics technology, will enable data centres to scale resources within or across server racks.
The unveiling of Photowave follows Lightelligence’s introduction of Hummingbird, an Optical Network-on-Chip (oNOC) processor for domain-specific artificial intelligence (AI) workloads back in June, and PACE, the world’s first integrated photonic computing system for Optical Multiply Accumulate (oMAC) in late 2021.
According to Yichen Shen, CEO of Lightelligence, “Photowave will soon set the standard for workload efficiency. It makes it possible to build scalable computing and accelerator pods with just the right amount of resources.”
As compute and memory stranding becomes an increasingly costly issue for data centre workloads, the industry-backed CXL standard enables the disaggregation of compute, accelerator, and memory resources to reduce stranding of resources while improving infrastructure efficiency. However, the current Ethernet-based infrastructure carries too much latency overhead to enable this disaggregation.
Photowave has been developed to provided a solution by interconnecting remote devices together using CXL over lower-latency fibre optic cable, extending reach to enable memory pooling at pod scales and beyond. This facilitates scalable CXL fabrics in the composable data centre.
The Photowave product line includes various form factors: a standard PCIe card, OCP 3.0 SFF card, and an active optical cable to achieve successful deployment of CXL-based infrastructure enhancements. They can be used in server platforms, CXL switches, memory appliances, and xPUs.
Photowave enables CXL 2.0/PCIe Gen 5 connectivity over optics with support for x16, x8, x4 and x2 bifurcation modes of operation, allowing for a wide variety of deployment scenarios. In addition to high-speed electrical-to-optical conversion for the data signals, Photowave is also able to support sideband signals over optics making it possible to implement more efficient and reliable disaggregation architectures.