Fully Homomorphic Encryption (FHE) has long been positioned as a breakthrough for secure computing. At its core, it allows organisations to compute on data without ever decrypting it, enabling searches, analytics, and AI workloads to run on external infrastructure without exposing sensitive information. The appeal is clear: FHE could unlock new forms of collaboration across healthcare, finance, government, and defence, as it enables computation across trust boundaries without compromising the privacy of individuals or the confidentiality of the data.

But the challenge has always been performance.

For security and malleability reasons, FHE introduces a significant data expansion. Ciphertexts are much larger than their plaintext equivalents, and the computations themselves rely on repeated execution of heavy mathematical transforms such as Number Theoretic Transforms (NTTs). The result is a double burden: more complex computation applied to much larger data structures. In practice, this leads not only to compute overhead, but to a substantial increase in data movement. 

The data movement bottleneck

This is where the story becomes more interesting.

As the industry begins to accelerate the compute side of FHE, a second bottleneck quickly emerges: moving data to and through the computation. Increasingly, it is not arithmetic that limits performance, but the cost of shuttling large, encrypted data between memory, processors, and interconnects.

Intel’s HERACLES project is a notable step forward in this context. The system combines specialised arithmetic units with a hierarchical memory architecture to accelerate core FHE operations while attempting to mitigate data movement overhead. It represents a highly optimised digital approach and arguably pushes digital-only architectures close to their practical limits for the current class of FHE workloads. 

However, even with such optimisations, the fundamental constraint remains. Data still needs to move through a digital system—across interfaces, memory hierarchies, and compute units. As FHE workloads scale, this movement increasingly dominates both latency and energy consumption. 

This is where alternatives emerge.

Photonic compute-in-transit: a novel architectural paradigm

Photonic computing, as developed by us at Optalysys, takes a fundamentally different approach. Instead of treating data movement and computation as separate steps, it merges them.

By performing arithmetic operations such as NTTs, inverse NTTs, additions and multiplications directly on optical data streams, computation occurs while the data is in transit.

This concept of “computing in transit” represents a shift in architecture. Instead of repeatedly moving data between memory and processing units, operations can be applied as data flows between components or even across network links. In effect, it collapses the traditional boundary between communication and computation.

The implications of photonic compute-in-transit

The implications are significant. Performing operations at optical line rates at hundreds of gigabits to terabits per second offers a path to reducing latency from microseconds to nanoseconds, while also lowering energy consumption per bit and easing thermal constraints in large-scale systems. More importantly, it directly targets the underlying issue in FHE: the cost of moving large volumes of encrypted data. 

The broader trend is clear. FHE is transitioning from a theoretical capability to an infrastructure problem. Early progress has focused on accelerating computation, but as those gains materialize, the bottleneck shifts to data movement. 

Intel HERACLES demonstrates how far digital acceleration can be pushed. Optalysys’ photonic approaches, meanwhile, suggest a different trajectory, one in which the architecture itself is rethought to address the root cause of the bottleneck. While practical deployment will depend on system integration, tooling, and ecosystem maturity, the architectural trajectory is becoming clearer. 

If FHE is to become practical at scale, the question may no longer be how fast we can compute, but how efficiently encrypted data moves and how much computation can be embedded within that movement.

---

At Optalysys we’re pioneering the architectural revolution enabling photonic compute-in-transit. Get in touch with us to find out how we can bring efficiency gains to your use case →