Dataflow 101
What Is an AI Chip?
How GPUs Process AI
How RDUs Process AI
Description
Learn what a Reconfigurable Dataflow Unit (RDU) is, how it differs from general-purpose AI hardware, and why it is designed specifically to make inference more efficient when trained models are serving real requests in production.
Additional Resources
Where GPUs Hit Limits
Key Differentiators of RDU for Inference
Description
An RDU, or Reconfigurable Dataflow Unit, is purpose-built for inference, designed around data locality, operator fusion, and continuous execution rather than raw compute. This video breaks down what makes RDUs different from GPUs and why that distinction matters for running AI models in production.
Additional Resources
- Blog: Why agentic inference needs hybrid hardware
- Blog: Why dataflow matters
-
Product: RDU
Understanding the Benefits of Model Bundling
Description
Discover how SambaNova's model bundling lets you run multiple AI models on a single rack, boosting cluster utilization and lowering total cost of ownership for agentic and multi-model workloads.
Additional Resources
- Blog: Why modern AI infrastructure needs model bundling
- Blog: How model bundling works
- Product: SambaRack
- Product: SambaStack