News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
SambaNova SN40L

SambaNova SN40L

AI inference chip with three-tier memory architecture

The SambaNova SN40L is a Reconfigurable Dataflow Unit designed for AI inference workloads. Announced in September 2023, it uses a unique dataflow architecture and three levels of memory to reduce data movement. This design aims to improve energy efficiency and throughput for large language models compared to traditional accelerators.

Share on WhatsApp

  • Announced in September 2023 for AI inference
  • Uses 5 nanometer TSMC manufacturing process
  • Features three-tier memory: SRAM, HBM3, DDR5
  • Clock speed of approximately 1.87 GHz
  • Predecessor to the SN50 Cerulean 2 chip

What it is

The SambaNova SN40L is a specialized processor for artificial intelligence inference. It belongs to the company's Reconfigurable Dataflow Unit family. The chip was announced in September 2023. It targets the memory-bound phase of running large models. Unlike general-purpose graphics processors, this hardware maps model execution directly onto the processor. This approach minimizes the need to move data back and forth to external memory. The goal is to achieve faster token generation and lower latency for interactive AI agents.

How it came about

SambaNova Systems developed the SN40L as part of its fifth-generation hardware strategy. The company announced the chip in September 2023. It was designed to address the high cost and latency of inference in generative AI. The architecture evolved from earlier Reconfigurable Dataflow Units. In November 2024, technical details were outlined in a paper. The chip serves as the predecessor to the SN50, which was later introduced to support larger agentic workloads with improved performance.

Inside the chip

The chip uses a Reconfigurable Dataflow Architecture. It chains operations into a single continuous flow rather than stopping at each step. This reduces overhead from relaunching kernels. The design includes Pattern Compute Units for math and Pattern Memory Units for storage. These units connect via a mesh interconnect. A key feature is the three-tier memory system. On-chip SRAM holds hot local data. High Bandwidth Memory stores active weights. DDR5 handles larger model pools. The SambaFlow compiler statically schedules data across these tiers to maximize locality.

How it was made

TSMC manufactured the SN40L using a 5 nanometer process node. The chip features 260 MB of SRAM per chiplet. It includes 64 GB of HBM3 memory with a bandwidth of 1.84 TB per second. The clock speed operates at approximately 1.87 GHz. The device supports up to 512 GB of DDR5 memory for prompt caches. This configuration balances compute density with memory capacity to handle concurrent workloads efficiently.

Where it went

The SN40L powers AI inference in data centers. It supports interactive agents, coding assistants, and conversational AI. The chip handles retrieval-augmented generation and multi-step reasoning tasks. It allows multiple models to run simultaneously by keeping them resident in memory. This capability helps enterprise copilots switch between models quickly. The hardware integrates into air-cooled or liquid-cooled racks. It suits environments where power constraints exist but low latency is critical.

Versions and family

The SN40L is the predecessor to the SambaNova SN50. The SN50, also known as Cerulean 2, offers higher performance and uses a 3 nanometer process. While the SN40L uses HBM3 memory, the newer SN50 reverts to HBM2E but doubles the core count. The SN40L remains suitable for existing deployments. The SN50 targets larger scale agentic workflows. Both chips share the same fundamental dataflow architecture and compiler stack.

Why it matters

The SN40L represents an alternative to GPU-dominated AI infrastructure. It demonstrates that dataflow architectures can compete in inference markets. By reducing data movement, it aims to lower the cost per token for generative AI. The chip supports the scaling of large language models without requiring new data center construction. Its design influences how memory hierarchy is managed in specialized AI accelerators.

Questions people ask about SambaNova SN40L

What is the SambaNova SN40L used for?

It is an AI inference chip designed for large language models and interactive agents. It handles tasks like coding assistance and conversational AI by reducing data movement latency.

How does the SN40L manage memory?

It uses a three-tier system. Fast SRAM sits on-chip for hot data. HBM3 stores active model weights. DDR5 holds larger prompt caches. This hierarchy keeps data close to compute units.

Who manufactures the SN40L chip?

TSMC manufactures the SN40L using a 5 nanometer process node. SambaNova Systems designs the architecture and compiles the software stack for the hardware.

What is the successor to the SN40L?

The SambaNova SN50, also called Cerulean 2, is the successor. It uses a 3 nanometer process and offers higher performance for agentic AI workloads compared to the SN40L.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on SambaNova SN40L yet, so the facts on this page come from the sources listed above: sambanova.ai/technology/sn40l-rdu-ai-chip, nextplatform.com/ai/2026/02/25/sambanova-pits-its-engineering-against-. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.