SambaNova

Overview

SambaNova Systems is an AI chip company that designs the Reconfigurable Dataflow Unit (RDU), a hardware accelerator purpose-built for AI inference with a fundamentally different architecture from GPU-based approaches. Rather than using fixed-function GPU cores, SambaNova's RDU uses a reconfigurable dataflow fabric that can be optimized for the specific patterns of AI workloads. The company's SM50 is its latest rack-level system, housing RDU accelerators alongside conventional compute.

SambaNova has positioned itself in disaggregated inference architectures, where heterogeneous compute resources — CPUs, GPUs, and RDUs — each handle the phases of an AI inference pipeline best suited to their architecture. In a joint demonstration with Intel at Computex 2026, a disaggregated stack using SambaNova SM50 RDUs for token decode, Intel Xeon 6 for tooling and orchestration, and Nvidia GPUs for prompt caching and prefill achieved 3.1x lower latency than a GPU-only stack.