Reconfigurable Dataflow Unit (RDU)

Delivering fast and energy-efficient inference

Introducing the SN50

Purpose-built for agentic inference, our fifth-generation chip, the SN50, is the only chip to deliver the speed and throughput required for agentic AI.

Built on the dataflow architecture, the SN50 delivers the best tokens per watt with 5X more compute and 4X more network bandwidth than our fourth-generation SN40L.

Learn more

Headline here

From chips to racks

The combination of 16 SN40L RDUs creates a single, high-performance rack that can run the largest models, such as DeepSeek R1 671B and Llama 4 Maverick, with the fast inference. These racks can be seamlessly integrated into any existing air-cooled data center.

Learn more →

Seamlessly achieve high performance

From chips to racks

The combination of 16 SN40L RDUs creates a single, high-performance rack that can run the largest models, such as DeepSeek R1 671B and Llama 4 Maverick, with fast inference. These racks can be integrated seamlessly into any existing air-cooled data center.

Learn more →

Solving AI’s data movement problem

Data movement is the most expensive operation when running AI. SambaNova RDUs are designed to solve this problem by using an architecture that creates an assembly line process on the RDU chip.

Our dataflow technology moves data seamlessly from operation to operation, saving power and time when processing the largest of models.

Learn more →

Tiered memory supports the largest models

RDUs employ a unique three-tier memory architecture that enables scaling to the largest LLMs. The infrastructure can scale to support running and switching between multiple models in milliseconds.

Learn more →

Find out how SambaNova is driving AI innovation

Chart - Gen speed VS Gen Throughput - Llama 3.3 70B - v2

The best speed and throughput in the Goldilocks Zone

SambaNova RDUs deliver low latency with high throughput, resulting in better tokenomics for use cases like AI coding agents that require near-real time inference.

Learn more →

Energy-efficient AI inference

Our dataflow architecture delivers extraordinary performance without the overhead of moving data back and forth between memory like GPUs.

Our fourth-generation chip, the SN40L, delivers fast inference with an average of just 10 kWh, allowing SambaRack systems to be air cooled.

Learn more →

2026 02 13 - Energy Efficient SN50 v4.2-07

From chips to racks

The combination of 16 RDUs creates a single platform that can run the largest models, up to 10 trillion parameters and 10 million tokens with our fifth-generation SN50 chips. These racks can be integrated seamlessly into any existing air-cooled data center.

Learn more →

Built for cloud scale

New with the SambaNova SN50 chips is a scaleout network of up to 32K RDUs. This enables huge cloud-scale inference services making it the ideal solution for inference service providers, like neo-cloud providers and hyperscalers.

Learn more →

Dataflow architecture

Our innovative compute and memory chip layout enables seamless dataflow between operations when processing AI models. This approach results in high-speed data traffic and significant gains in performance and efficiency.

Learn more →

Three-tier memory for efficiency

The SN40L design enables multiple models to run in memory and switch models in microseconds. This unique layout enables SambaNova to scale to the largest models, like DeepSeek and Llama 4 — all on a single rack.

Learn more →