Reconfigurable Dataflow Unit (RDU)

GPUs are general-purpose accelerators designed to handle a wide range of compute workloads, primarily for training. The SambaNova RDU is purpose-built for inference and uses Dataflow architecture and three-tier memory architecture that maps model execution directly onto the processor, minimizing data movement to memory, which is the most expensive component for AI inference.

The SN50 is the latest generation of SambaNova’s RDU, offering higher compute performance, increased network bandwidth, and improved scalability compared to the SN40L. While the SN40L is well-suited for existing inference deployments and power-constrained environments, the SN50L is designed for large-scale, agentic AI workloads. It enables faster token generation, better system throughput, and more efficient multi-model execution.

The SN50 supports a wide range of inference-heavy AI workloads that require low latency, high throughput, and efficient memory usage. These include AI agents, coding assistants, enterprise copilots, conversational AI, retrieval-augmented generation (RAG), and model hosting platforms. It is particularly well-suited to agentic workflows involving multi-step reasoning, tool usage, and frequent model switching.

Yes, the SN50 is designed to run multiple models simultaneously using its tiered memory architecture. This allows models to remain resident in memory and be switched quickly with minimal latency. The capability is especially important for agentic workloads that rely on multiple models across task steps, improving responsiveness, utilization, and overall inference efficiency.

The SN50 scales to large models through a combination of memory architecture and distributed deployment. Multiple racks can be interconnected to form inference clusters, enabling support for larger models, higher concurrency, and predictable performance at scale.