Agentic AI changes the demands placed on AI infrastructure. Instead of generating a single response, an agent may reason through multiple stages, retrieve information, use tools and repeatedly call a model before completing a task.

In this presentation, Pushkar Nandkar explores SambaNova's SN40L and the dataflow architecture underlying its approach to AI inference.

SambaNova's Reconfigurable Dataflow Unit (RDU) is designed differently from conventional GPU architecture. By optimising computation around the flow of data through the model, the architecture is designed to deliver fast inference while supporting demanding large-model workloads.

For agentic applications, inference speed matters because delays can accumulate across multiple model calls. Improving the performance of each inference step can therefore reduce the time required to complete the overall workflow.

The presentation demonstrates why the infrastructure beneath an AI model becomes increasingly important as organisations move from conversational AI towards production agentic systems.