Products
Technology
Resources
Solutions
Community
About

The fourth generation RDU powering the SambaNova Suite

SambaNova delivers the only true full stack platform, purpose built for generative AI. The SN40L is at the heart of SambaNova platform and uses a revolutionary design to offer both dense and sparse compute and to include both large and fast memory, making it a truly “intelligent chip”.

The SN40L is easily able to perform both training and inference on a single platform, eliminating the redundancy of legacy systems and to power the largest, multi-trillion parameter models that are enabling enterprise-grade generative AI.

Dataflow Architecture

The SN40L has a reconfigurable dataflow architecture which is automatically programmed and optimized to best suit how data should be processed across the chip, dramatically improving the hardware utilization and memory consumption when compared to legacy GPU technology.

Dense & Sparse Compute

The SN40L has a reconfigurable dataflow architecture makes it capable of simultaneously processing both sparse data structures, that consist of mostly zero values, as well as very dense data structures. This enables high utilization of the underlying platform while allowing a diverse set of models to be easily utilized.

Three-tiered memory

Key to the SN40L is a new three-tiered memory architecture that includes both very fast memory, high bandwidth, and very large memory. This enables the system to handle the largest, multi-trillion parameter models of today and tomorrow, without the costly parallelization required with legacy, GPU-based systems. 

Training & Inference

The combination of dense and sparse compute with the three-tiered memory architecture makes it possible to efficiently perform both training and inference on a single platform, eliminating the need for multiple systems for these functions with the largest models.

Ready to see more?

Schedule a meeting to see how SambaNova can help advance your AI initiatives.