Summary
- iPronics ONE will be integrated into BSC's research GPU and HPC infrastructure.
- The collaboration will test workload-aware optical networking across training, inference, LLM, and mixture-of-experts workloads.
- The programme targets network utilisation, latency, energy consumption, and GPU efficiency rather than additional data centre floor space.
iPronics and Barcelona Supercomputing Center have begun a two-year programme to integrate programmable optical circuit switching into GPU and high-performance computing infrastructure.
The collaboration will place iPronics ONE inside BSC’s research environment and combine the optical hardware with software developed above the switch-management layer. The teams intend to test workload-aware networking across AI training, inference, large language models, and mixture-of-experts workloads.
The work addresses a constraint that becomes increasingly physical as accelerator clusters grow. Adding GPUs only increases useful computing capacity if those processors can exchange data quickly enough for the workload they are running.
Conventional data centre networks rely heavily on electronic packet switching. Optical circuit switching offers a different tool: establishing configurable optical paths that can be changed according to where traffic needs to move rather than forcing every flow through a permanently fixed topology.
For AI facilities, the attraction lies in potentially reducing network bottlenecks and making expensive accelerators spend less time waiting for communication.
Network utilisation becomes facility economics
GPU utilisation has a direct infrastructure consequence. A processor that is powered and cooled but underused because the network cannot feed it effectively still consumes part of the data centre’s electrical and thermal budget.
That makes networking efficiency relevant beyond the IT layer. If an operator can extract more useful computation from the same installed GPU estate, the return on the power connection, cooling plant, racks, and building supporting those machines improves as well.
The BSC collaboration is designed to examine that relationship under research conditions rather than simply demonstrating an isolated optical component.
iPronics will provide the optical switching platform, APIs, and lower-level software. BSC will develop higher-level controls intended to align network behaviour with workload requirements.
That software element is critical. A reconfigurable fabric only becomes useful if orchestration can understand when a topology should change and do so without creating instability or unacceptable delays.
Optics move deeper into AI systems
Optical technology is already fundamental to data centre connectivity, particularly between switches and across longer distances. The current wave of development is pushing optics closer to the compute and making the optical layer more programmable.
That trend is being driven by electrical limits. As bandwidth rises, moving data through copper over even relatively short distances becomes harder to achieve within acceptable power and thermal budgets.
Optical circuit switching does not remove every electronic switch, nor does it automatically suit every traffic pattern. AI workloads can produce complex communication behaviour, and fixed packet networks remain extremely flexible for diverse traffic.
The question for the BSC programme is where a dynamically reconfigurable optical layer delivers enough benefit to justify additional control complexity.
The collaboration is particularly useful because BSC provides an operating HPC research environment rather than a laboratory connector demonstration. Results can therefore expose practical issues around orchestration, failure recovery, workload scheduling, and integration with existing software stacks.
If programmable optics improve GPU utilisation or reduce the network’s energy cost, the effect ultimately appears in the economics of the complete data centre. High-density AI facilities are increasingly constrained not just by how much compute can fit into a rack, but by how effectively power, cooling, and connectivity allow that compute to work as one system.
