PsiQuantum Aims to Get More Useful AI Work From the Network

A room full of powerful processors can still spend valuable time waiting. When a computing job is divided among machines, one part may need another part's results before it can continue. A congested connection can hold up the next calculation even though the chip assigned to it is ready. Buying faster processors does little for that particular delay. 5

That is the opening for a device PsiQuantum unveiled on October 28: an optical circuit switch (OCS) that changes connections while keeping the traffic as light. The company reports laboratory measurements on an 8×8 demonstration device and is sharing samples with selected prospective customers. It is applying photonics developed for its quantum-computing program to artificial intelligence (AI) infrastructure. The announcement concerns conventional data networks, rather than a new quantum computation. 1

The potential gain has two connected parts. Removing avoidable electronic switching stages could reduce the electricity and equipment devoted to moving data. Giving busy groups of machines better connections could also reduce waiting, allowing installed processors to finish more useful work. Both depend on the network design and workload. PsiQuantum has not demonstrated a fleet-wide gain in computing output per unit of electricity.

Keeping a connection in light

Data commonly travels through fiber as optical signals. Where an electronic packet switch handles that traffic, a receiver turns the arriving signal into electrical data; switching electronics choose its onward path; a transmitter sends it back into fiber. A packet is a small addressed piece of a larger exchange. This arrangement gives networks flexible decisions about many simultaneous conversations, but the interfaces and switching stages consume power. 6

An OCS instead establishes a connection from one optical port to another. The light passes through that connection without being converted into electrical data inside the switch. The endpoints still need transmitters and receivers, and the wider network still needs control electronics and ways to direct traffic. The opportunity is to remove selected intermediate conversions and switching stages from a route, rather than eliminate electricity from networking. 2

Consider two groups of servers that repeatedly exchange large amounts of data. A suitably designed optical layer could connect them more directly instead of sending those exchanges through extra electronic switching stages. That connection would remain available while the traffic pattern justified it, then could be reassigned. This is an illustrative use, not a measured PsiQuantum customer deployment.

Connections are scarce resources. Dedicating more of them to one busy pair leaves fewer for other pairs, and real jobs do not all communicate in the same way. A controller must decide which connections to establish, when to change them and how traffic should travel during the change. A fast component can therefore fail to improve a poorly matched network. The design has to serve changing demand as well as a favorable demonstration.

PsiQuantum describes its device as having no moving parts and being strictly non-blocking: an available input-output connection can be established without disturbing other established connections. Its typical insertion loss is 1.2 decibels (dB), a logarithmic measure of optical attenuation. Lower loss preserves more of the incoming optical power for the receiver and leaves more allowance for fibers, connectors and other components along the route. That is an optical budget, not a measurement of electrical energy saved. 2

The reported reconfiguration time is under three milliseconds. This measures changing a connection, not the time each packet takes to cross an already established path. Treating it as packet latency would confuse rearranging the network with carrying traffic over it. Whether such a change is sufficiently quick depends on how long the new connection will be useful and how the system handles the transition. 1

Google has already deployed the architectural idea

Optical circuit switching has an established data-center precedent. Google's 2022 Jupiter paper describes a production network combining optical switches, software control and more direct connections between groups of machines. Its switches used micro-electromechanical systems (MEMS), tiny moving structures; PsiQuantum's demonstration uses a different implementation. The precedent establishes that an optical switching layer can be operated at scale, not that every implementation will deliver the same outcome. 3

Google reported a 41 percent reduction in network power and a 30 percent reduction in network capital costs across the architectural evolution described in that paper. Removing a layer of electronic switches and associated optics accounted for much of the power reduction. Those are results for Google's network architecture and comparison baseline. They are neither whole-data-center savings nor a forecast for PsiQuantum's switch. 3

This boundary is essential to understanding the electricity opportunity. A facility also powers processors, memory, storage, cooling and supporting equipment. Reducing one portion of its network demand changes only part of the total. Moreover, if improved connections let processors spend less time waiting, their power use may rise as they do more work. An operator should measure completed work against total energy, not assume that better utilization necessarily lowers the electricity meter.

There are consequently two useful outcomes to distinguish. One is finishing a fixed workload with less energy. The other is completing more work within an existing power limit. Either could make installed infrastructure more productive, but they require different measurements. A component's low optical loss alone establishes neither result, and a network improvement cannot be translated into additional useful compute without examining what was limiting the job.

From a small switch to a useful system

An 8×8 switch provides eight input and eight output ports. It is a small hardware building block, not a complete interconnection network for a large computing cluster. Increasing port count, packaging many optical connections and maintaining consistent behavior across manufactured devices introduce further engineering demands. Customer samples can expose those demands under conditions a supplier's laboratory may not capture.

The disclosed result also leaves an economic question open. Fewer electronic stages might save equipment and energy, while optical components, packaging, control and maintenance add their own costs. A replacement has to be compared with the network it would actually displace over a useful service life. PsiQuantum has not supplied customer-fleet evidence establishing that balance for this device.

For a buyer, the revealing demonstration would connect the hardware result to an application: run a representative job over a defined network, count completed work, measure total energy and record what happens as demand changes or a connection fails. The comparison would include the optical endpoints and control system. That would show whether a cleaner data path actually releases capacity the installed processors can use.

The broader opportunity is to improve the work a data center obtains from equipment it has already bought and electricity it can already secure. PsiQuantum's laboratory switch offers a possible component for that effort. Its next milestones are larger, reliable devices and customer evaluations that connect optical performance to useful computing work. The architectural case is established; the contribution and economics of this particular implementation remain to be demonstrated.

Sources

• PsiQuantum October 28 announcement and laboratory scope

• PsiQuantum optical circuit switching specifications

• Google Jupiter Evolving paper and production-network comparison

• Google Jupiter Evolving full paper

• NVIDIA guidance on distributed training communication bottlenecks

• Ho and colleagues on optical interconnects and their electrical interfaces