Optalysys aims to compute as data travels, not just when it arrives

Optalysys has demonstrated photonic hardware capable of performing tens of GFLOPs of operations directly on data as it travels, a departure from conventional systems where data must first reach processing units, the company says. The company’s approach tackles a growing bottleneck in computing: data movement, which increasingly limits performance for demanding workloads like artificial intelligence and cryptography. This compute-in-transit model shifts the focus from simply moving data to applying computation as it propagates through the system.

Data Movement Bottlenecks in AI and Cryptography

This compute-in-transit model addresses a critical limitation in current systems; data movement is increasingly the dominant bottleneck, particularly for demanding applications like artificial intelligence, Post-Quantum Cryptography, and Fully Homomorphic Encryption. Conventional architectures rely on transporting data to discrete processing units, creating inefficiencies in latency, bandwidth, and energy consumption as system scale increases. AI systems currently depend heavily on multiply-accumulate operations on floating-point data, a computationally intensive process; however, cryptographic workloads such as Fully Homomorphic Encryption utilize structured polynomial transformations, including Fast Fourier Transforms and Number Theoretic Transforms, as their computational foundation.

Although differing in structure, both areas involve repeated application of regular operations over large datasets, making performance increasingly sensitive to data movement and memory access patterns. The creation and propagation of intermediate data further exacerbate this issue, requiring repeated storage, transfer, and transformation across processing stages.

Digital Signal Processing cores and Application-Specific Integrated Circuits accelerate common operations, but scaling to very large transform sizes, like the 65,536-point NTTs needed for certain Fully Homomorphic Encryption workloads, remains constrained by memory access and interconnect complexity. Optalysys states that their prototype uses a Photonic Integrated Circuit to apply computation directly along the data path during transmission, operating through a standard QSFP56 port. The prototype includes self-link fault detection and automated photonic calibration mechanisms, achieving sustained operation.

Photonic Integrated Circuits Enable Programmable Compute-in-Transit

Current computer designs separate data transmission from processing, creating a performance bottleneck as workloads grow more demanding. A prototype photonic integrated circuit supports a range of mathematical operations through a programmable serial compute pipeline operating across high-speed transceiver lanes. This tighter integration between computation and data movement provides a foundation for architectures that more closely align computation with dataflow.