Salience’s $125 million round backs an AI network that switches light

Salience Labs’ $125 million Series B is funding a route into the equipment between AI processors. Announced on 14 October, the round was led by Bullhound Capital, with AMD Ventures and Intel Capital among the participants. The Oxford company plans to expand manufacturing, deployment and its Silicon Valley presence.[1]

Its existing OC-32M module connects 32 input fibres with 32 output fibres through a non-blocking optical switch. The module combines a photonic chip for switching and amplification with driving electronics, firmware and fibre connectors. Salience also offers an evaluation platform for customers developing systems around it.[2]

The distinction is where the signal changes form. Salience’s all-optical design removes the conversion into electronics at the switch that its comparison calls optical–electrical–optical, or OEO. The company presents that as a way to reduce switching power and equipment cost. These are claims about its network architecture, rather than a disclosed saving on a customer’s entire AI operation.[3]

An optical switch has two clocks

The OC-32M product page specifies 10 nanoseconds for travel from an input port to an output port, and less than 300 microseconds to reconfigure the connections. It also specifies less than one watt per port. The timing figures describe different operations: crossing an established route and setting up a changed route.[2]

The implication is that scheduling becomes part of the value proposition. A very short journey through the switch can coexist with a much longer pause while a new path is selected. A workload that keeps useful connections busy can spread that setup time across more transferred data. Repeatedly changing paths for tiny transfers would create a different balance.[3]

For an illustrative calculation, assume a 0.3-millisecond setup followed by 10 milliseconds of useful transfer, with no overlap. Setup occupies about 2.9% of the combined 10.3 milliseconds. If the transfer lasts only 0.1 milliseconds, the same setup occupies 75% of the total. We chose the setup value to illustrate the published reconfiguration scale; these are hypothetical transfers, not measured Salience workload results.

Port-to-port latency answers how quickly an established connection carries a signal. The pattern of connection changes determines how often the system has to pay for another setup. An operator needs both to understand where the component fits.[3]

From a demonstration to a manufactured product

Salience and Keysight announced a joint testing environment in March, extending Keysight’s KAI Data Center Builder validation capabilities to optical circuit switching. Their programme covers how the interconnect affects training and inference. That partnership provides a route for examining the assembled network, beyond the individual switch’s specifications.[4]

The latest funding follows Salience’s demonstration of a 64-port switch at ECOC. The company names 128- and 256-port versions as future developments, and says it will continue working with partners including Tower Semiconductor and Keysight as manufacturing scales. The demonstration, current 32-port module and planned larger versions occupy different stages of that programme.[1]

For investors, the commercial proposition is a component that other companies can build into AI systems. More optical paths need control, integration and a repeatable supply of hardware. The round backs making that component available in volume, where a rack builder can evaluate it within a complete machine.[1]

Report an error or suggest a correction