Description du poste
The OpportunityLumai is moving from optical-compute research into production AI systems, and the architecture and compiler decisions that shape the product are being made now, ahead of the silicon that will confirm them. This hire builds the modelling capability those decisions rest on.
The RoleYou will own the performance and power modelling framework that predicts how Lumai's photonic AI hardware performs against real inference workloads, giving the architecture and compiler teams one shared source of truth for latency, throughput and power across device configurations. You will model the compute engines, interconnect, memory and bridge interfaces; keep the model honest by characterising operator-level workloads against hardware, RTL or emulation results; and turn what the model says into decisions teams act on.
What is genuinely hard about this is balancing speed, fidelity and usability. The model has to capture concurrency, event-driven behaviour and compute-block latency in enough detail to support real architecture decisions, and still compare configurations in minutes rather than overnight. The gap between what the model predicts and what the hardware does has to be measured rather than assumed, and closing it is the part that never stops.
What You'll DoDesign, build and maintain a Python-based framework that models hardware and software latency, hardware concurrency and event-driven relationships.
Model compute engines, interconnect and synchronisation fabric, memory