RDU: The GPU Alternative

Power the most demanding generative and agentic AI workloads with the most performant and capable processor, purpose-built for AI

Built for Agentic AI

With TBs of addressable memory, the SN40L reconfigurable dataflow unit (RDU) was built for the next generation of AI workloads, known as Agentic AI. Agentic AI connects multiple models to work together to perform complex tasks that individual models cannot do alone, all from a single user prompt.  This is a complex task that cannot be effectively performed with GPUs. 

World record inference performance

Inference is rapidly replacing model training as the primary function of AI systems. The combination of the large addressable memory and the dataflow architecture of the RDU result in a system that is significantly faster than other processors for model inference, shown by the multiple world records. While other processors can only perform inference, the SambaNova RDU is capable of model training and inference on a single system. With the SambaNova RDU you can train and fine tune your models on the same platform you run inference, all with the same or better training performance than GPUs.

Dataflow architecture

Designed for AI, the SambaNova RDU was built with a revolutionary dataflow architecture. This design makes the RDU significantly more efficient for these workloads than GPUs as it eliminates redundant calls to memory, which are an inherent limitation of how GPUs function. This built-in efficiency is one of the features that makes the RDU capable of much higher performance than GPUs in a fraction of the footprint.

Built for the AI workloads of today and tomorrow

Protect your investment in AI infrastructure with the only complete platform that can efficiently run the generative AI workloads of today and the agentic AI solutions of tomorrow. The SambaNova systems are capable of seamlessly scaling out, so it can grow with your organization as you increase your AI capabilities.