SambaNova RDU
Reconfigurable Dataflow Unit accelerators designed to run large models at low latency; the current chip is SN50, SambaNova's fifth-generation RDU.
Find alternatives to SambaNova RDU
What it does
- Accelerator silicon
- Model inference
Features 3
- Dataflow executionChains operations into a continuous dataflow across the processor to reduce memory transfers and overhead.
- Three-tier memoryOn-chip SRAM, HBM2E and DDR5 place data according to access speed and frequency.
- Millisecond model hot-swapModels residing in HBM and SRAM can be hot swapped in milliseconds, for agentic workloads that switch between models.
How it compares
-
Cerebras WSE-3 Turbo (WSE-3T)
Cerebras uses wafer-scale processors to deliver high-speed inference through its cloud, dedicated systems, and on-premise deployments.
Research report · 19 Sep 2026 -
NVIDIA Blackwell
NVIDIA dominates this space with roughly 85% market share through their H100 GPUs and CUDA software ecosystem.
Research report · 19 Sep 2026
Sources
- Pricing
-
Official documentation · 21 Aug 2026
Sold inside SambaNova systems rather than as a listed part; no unit price is published.
- Description
-
Official documentation · 22 Oct 2026
RDU page, read 2026-10-22: "SN50 is SambaNova’s fifth-generation Reconfigurable Dataflow Unit (RDU) chip, purpose-built hardware for the memory-bound phase of inference."
- What it does
-
Official documentation · 7 Sep 2026
Fetched 2026-09-07: HTTP 200, <title> “RDU | Next-Gen AI Chip for Inference at Scale”. “The SambaNova RDU is purpose-built for inference”. TRAINING IS DELIBERATELY NOT ADDED: the page never claims it, and its own FAQ distinguishes the RDU from GPUs, which are “primarily for training”. The absence is the vendor's, not the record's.
- Sold within
-
The RDU page says "Combining RDU chips into racks powers the largest models with the fast inference." and its only route is "Talk to an expert"; the SambaRack page says it "combines 16 SN50 RDU chips in one rack". No page offers the chip on its own or states a price.
- Deployment
-
Official documentation · 18 Oct 2026
Page: 'SambaRack systems can be easily integrated into both air-cooled and liquid-cooled data centers'.
Also from SambaNova
-
SambaCloud Model API
Hosted inference service serving open-weight models on SambaNova's own RDU hardware, billed per token.
-
SambaStack Infrastructure service
Full-stack AI system pairing SambaNova's RDU hardware with its serving software, deployed in a customer's own data centre.
-
SambaRack Hardware
Air-cooled, rack-scale AI inference hardware system integrating 16 SN50 RDU chips per rack (scalable across multiple racks); accessed via on-premises deployment, SambaCloud, or contact-sales for enterprise/government deployments.
Open SambaNova in the directory
Something wrong here? Send a correction — quote this product id: sambanova-rdu.