SambaNova's Reconfigurable Dataflow Unit (RDU) is a dataflow architecture chip family built for enterprise AI training and inference. The SN40L generation introduced a single-card design combining compute with high-capacity memory to run large models without external DRAM bottlenecks. The following SN50 generation continues the dataflow approach with larger on-chip memory and higher compute density for generative and agentic AI workloads.

SambaNova positions RDU-based systems for inference serving, including agentic workflows that require low latency across multiple model calls. The chips are packaged in the company's DataScale and Rack Scale systems, marketed as alternatives to GPU-centric AI infrastructure for private-cloud and on-premise deployments.