Flumen: Dynamic Processing in the Photonic Interconnect

In chiplet-based heterogeneous architectures, electrical network-on-package (NoP) designs are typically over-provisioned with routers and channels to provide sufficient bandwidth during periods of high network load. Observing that there are significant periods of low/idle network utilization, prior work has proposed modified network-on-chip (NoC) architectures to enable in-network compute, especially for compute-intensive operations (e.g. linear algebra). However, electrical package-level interconnects impose fundamental energy and bandwidth scaling issues for future chiplet architectures. This paper proposes Flumen, a dual-purpose photonic interconnect that provides communication at the package-level while also doubling as an accelerator, performing parallel linear computation when network load is low. The proposed architecture utilizes the inherent parallelism of light to create energy-efficient interconnects that support en route computation with minimal changes to the network. By dynamically adjusting the topology, Flumen can change the communication and compute sections of the architecture to adapt to workload fluctuations. Performance evaluation on linear algebra applications shows that Flumen achieves a 2.5x reduction in energy, a 3.6x speedup improvement, and a 9.3x reduction in energy-delay product on average when compared to an electrical mesh network that is used exclusively for communication.