Salience Labs develops silicon photonics optical switch technology for GPU interconnect, offering higher bandwidth and lower latency vs. NVIDIA NVLink copper.

Optical circuit switches have been used in network research labs and telecom production for over two decades, but they are now poised to move into the scale-up networking domain of high-performance computing clusters. Salience Labs has developed an optical circuit switch based entirely on silicon photonics technology, derived from advanced research aimed at creating a photonics computing platform.

Two types of optical switches are currently in production. MEMS-based switches use arrays of tiny mechanical mirrors that spin to create fiber-optic circuits, while liquid crystal on silicon (LCoS) switches rely on a different optical mechanism. Google has used MEMS devices called "Palomar" as part of its "Apollo" OCS backbone since 2015, enabling clusters to be reconfigured dynamically and supporting up to 9,216 TPUs in coherent memory clusters. Lumentum sells MEMS-based OCS equipment and has partnered with Nvidia, while Coherent provides LCoS-based OCS devices, also partnering with Nvidia.

Salience Labs has not disclosed its specific switching mechanism, but it uses neither MEMS nor LCoS. The company spun out of Oxford University and the University of Münster, founded on the research of Harish Bhaskaran, a professor of applied nanomaterials at Oxford, and Wolfram Pernice, a professor of experimental physics at Münster—both co-founders at Salience Labs. Their research focused on phase change optoelectronics. The company's partnership with Tower Semiconductor for PH18DA integrated III-V lasers and TPS45PH low-loss silicon nitride waveguides suggests it may be using phase change technology within a silicon photonics framework.

CEO and co-founder Vaysh Kewada—who holds a bachelor's and master's degree in physics from Imperial College London and served as an entrepreneur in residence at Oxford Science Enterprises—declined to specify the exact switching mechanism. The company is rolling out a 32-port OCS and plans to deliver 64, 128, and 256-port systems for scale-up AI networks.

What distinguishes Salience Labs' approach is speed. MEMS and LCS switches require milliseconds to reconfigure port-to-port links, while silicon photonics waveguide switching operates in under 300 microseconds. "We chose to develop OCS because there was quite a large number of players developing solutions on the CPO, NPO, XPO front, trying to solve how to get data off the chip and onto optical fiber," Kewada said. "When there were more optical connections established in the datacenter, we thought about what the switching architecture would look like. We developed our products with a simple hypothesis: there would be more optical connections in the datacenter, and when that happened, the switching layers would move towards more heterogeneous architectures—combinations of electrical packet switches and optical circuit switches. Our view is that OCS was a market ripe for disruption because the OCSes available today are based on technology that is twenty years old if not older."

The Salience Labs OCS comprises two chips: a fully integrated silicon photonics OCS and an amplification and signal conditioning chip, similar to retimers and redrivers for high-bandwidth copper circuits. Mounted back-to-back on a PCB card, the design solves a key silicon photonics challenge—optical loss. "As soon as you integrate on chip, you incur optical loss," Kewada explained. "We solved that using amplification with our own component design, fabricated on arrays to meet bill-of-materials cost goals. This amplification chip is one of a kind and the crux of the stack."

The 32-port system that launched in March operates at 100 Gb/sec native line rate with PAM4 modulation, achieving 200 Gb/sec effective bandwidth per lane. The architecture scales to 64, 128, and 256 ports. Port-to-port latency on the OC-32M is under 10 nanoseconds—main memory speed—compared to around 250 nanoseconds for the Broadcom Tomahawk Ultra Ethernet ASIC and 450-650 nanoseconds for other high-performance switches. This represents 25X lower latency than Tomahawk Ultra.

Reconfiguration time is dramatically faster than MEMS-based OCS by a factor of 3X or greater. Energy consumption per port is 8X lower for either OCS type compared to fast Ethernet switches. Potential customers are interested in expanding the scale-up domain for AI systems, which is currently limited to 72 GPUs from Nvidia and AMD. Some seek to extend coherent memory domains beyond single racks, while others want alternatives to existing rackscale memory fabrics or need lower-latency solutions for connecting multiple machines in the decode phase of generative AI. Salience Labs' OCS can package up to eight modules into a single 1U chassis.