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
NorthPole

NorthPole

Brain-inspired AI inference chip by IBM Research

NorthPole is a digital AI chip for neural inference designed by IBM Research. Built on a 12 nm process, it integrates memory and processing to bypass the von Neumann bottleneck. The chip offers high energy efficiency and low latency for tasks like computer vision.

Share on WhatsApp

  • Contains 22 billion transistors on a 12 nm process
  • 25 times more energy efficient than 12 nm GPUs
  • Integrates all memory on-chip to remove bottlenecks
  • Designed specifically for AI inference tasks
  • Operates with simple air cooling instead of liquid

What it is

NorthPole is a specialised digital chip created for artificial intelligence inference. IBM Research developed the device to address the inefficiencies of traditional hardware when running modern AI systems. Unlike standard processors that separate memory from computing units, this chip keeps all memory on the die itself. This design choice removes the data transfer delays known as the von Neumann bottleneck. The result is a system that processes information with significantly lower latency and higher energy efficiency than conventional graphics or central processing units.

How it came about

The project represents nearly two decades of work by Dharmendra Modha and his colleagues at IBM. It builds upon earlier research into brain-inspired computing, specifically following the TrueNorth chip released before 2014. For eight years, the team worked in secrecy at the Almaden lab without publishing papers. They aimed to combine the structural efficiency of biological brains with the speed of silicon. The first results were published in Science, marking a shift from theoretical models to a functional prototype capable of real-world tasks.

Inside the chip

The architecture blurs the line between storage and calculation. At the core level, it acts as memory-near-compute, while appearing as active memory from the outside. It contains 256 cores, each performing 2,048 operations per cycle at 8-bit precision. This count can double or quadruple with lower precision settings. Because it relies entirely on onboard memory, it cannot easily access external data. To handle larger networks, the system uses a scale-out approach, breaking models into sub-networks across multiple chips. It is designed strictly for inference, not training.

How it was made

The chip was fabricated using a 12 nanometre process technology. It packs 22 billion transistors into an area of 800 square millimetres. This density allows it to outperform older 14 nm CPUs and 12 nm GPUs in efficiency metrics. The design does not require complex liquid cooling systems. Standard fans and heat sinks are sufficient for thermal management, which enables deployment in smaller physical spaces where bulky cooling infrastructure is not available.

Where it went

Initial testing focused on computer vision tasks such as object detection and image segmentation, partly due to funding from the U.S. Department of Defense. It has also shown capability in natural language processing using encoder-only models and speech recognition. Potential applications include autonomous vehicles handling rare edge cases, robotics, and satellite monitoring for agriculture. The chip suits edge applications that need real-time data processing without heavy power demands.

Versions and family

NorthPole is the successor to the TrueNorth chip, which was inspired by bee-brain synapse counts. While TrueNorth was fully neuromorphic, NorthPole merges traditional processing speeds with brain-like structure. There are no commercial variants listed yet. The current design is a prototype intended to demonstrate architectural possibilities. Future iterations may leverage smaller manufacturing nodes, such as 3 nm or 2 nm technologies, to further improve performance.

Why it matters

The chip demonstrates that integrating memory and compute can drastically reduce energy use and latency. It challenges the dominance of traditional von Neumann architectures in AI workloads. By achieving high efficiency on a mature 12 nm node, it suggests that architectural innovation can outweigh raw process scaling. This could allow AI deployment in resource-constrained environments like remote sensors or mobile robotics, reducing the cost and complexity of large-scale AI infrastructure.

Problems and reception

The research team maintained a hermetic development process, avoiding public lectures until the publication of their findings in Science. The company claims the chip outperforms major prevalent architectures, including those built on more advanced 4 nm processes, in specific efficiency metrics. The reception highlights its potential for enterprise models rather than large language models like GPT-4, positioning it as a specialised tool for specific industrial and defence applications.

Questions people ask about NorthPole

What makes NorthPole different from standard chips?

Standard chips separate memory and processing, causing data transfer delays. NorthPole keeps all memory on the chip itself, removing this bottleneck. This integration allows for much faster and more energy-efficient AI inference compared to traditional designs.

Can NorthPole run large language models like GPT-4?

No, the chip is not designed for such large models. It relies on onboard memory, which limits its capacity. It is best suited for smaller enterprise models, computer vision tasks, and specific inference workloads that fit within its memory constraints.

How does NorthPole handle larger neural networks?

It uses a scale-out approach. Large networks are broken into smaller sub-networks that fit into the memory of individual chips. These chips are then connected together to process the full model, allowing the system to scale beyond single-chip limits.

What manufacturing process was used for NorthPole?

The chip was built using a 12 nanometre node process. Despite not using the most advanced nodes like 3 nm, its architectural efficiency allows it to outperform newer chips in terms of energy use and space efficiency for specific AI tasks.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on NorthPole yet, so the facts on this page come from the sources listed above: research.ibm.com/blog/northpole-ibm-ai-chip, tweaktown.com/news/93966/ibm-unveils-northpole-ai-chip-is-4000x-faster. 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.

    Photo credits on this page

    The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

    The weekly digest

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