NorthPole by IBM: status and timeline

NorthPole

A brain-inspired neuromorphic chip for AI inference with all memory on-chip, eliminating the energy cost of moving data between processor and memory. Delivers 25x better energy efficiency than leading GPUs.

ResearchStage 1 of 5

Research prototype (October 2023). 22B-transistor inference engine for efficient edge AI.

Updated 19 Oct 2023Checked 10 Oct0 updates this week

Milestones

No announced next step

NorthPole research prototype completed2023Complete.
Results published in ScienceOct 2023Complete.

Most important updates

IBM announces NorthPole brain-inspired inference chip with 25x better energy efficiency than GPUs19 Oct 2023

Current obstacles

Limited model size due to on-chip memory constraintsAll model weights must fit in on-chip SRAM; large models require off-chip memory or pruning, reducing the efficiency advantage.

Physics limits

On-chip SRAM is expensive in area and powerSRAM scales poorly; doubling the memory capacity roughly doubles area and leakage power, setting a practical limit on model size regardless of compute speed.
Inference-only use caseNorthPole cannot train neural networks, limiting it to edge and deployment scenarios; cloud training remains on GPUs and TPUs.

How it works

3 parts

On-Chip Memory

All memory integrated on processor

Memory sits where computation happens, eliminating slow bus traffic and achieving 13 TB/s bandwidth.

Neuromorphic Architecture

Brain-inspired cores

256 cores process data in parallel with local memory access, avoiding the energy cost of separating compute from storage.

Inference Only

Deployment-focused design

NorthPole runs trained models at low latency and power, not designed for training neural networks.

Spec sheet

Spec	NorthPole
Transistor count	22 billion^R (reported)
Core count	256^R (reported)
Operations per core per cycle	2,048 at 8-bit^R (reported)
Process node	12 nm^R (reported)
Die size	800 mm²^R (reported)
On-chip memory bandwidth	13 TB/s^R (reported)
Energy efficiency vs GPU	25x better (ResNet-50)^R (reported)

^R reported by the company

Update log

1 update

Thu 19 Oct 2023

Major: IBM announces NorthPole brain-inspired inference chip with 25x better energy efficiency than GPUsPressHardware

About IBM

The team behind NorthPole

IBM

Superconducting chips, error correction, Qiskit

IBM has run quantum computers on the cloud since 2016 and runs the largest fleet. It aims for Starling, an error-corrected machine with 200 logical qubits, by 2029. Its Qiskit software is the most used quantum toolkit.

Unveiled Nighthawk (120 qubits) and Loon (error-correction test chip) in November 2025.
Expects partners to show verified quantum advantage by the end of 2026.
Committed more than $10B to quantum computing in June 2026.

ibm.com (opens ibm.com)Company profile →

Founded: 1911115 yrs
Headquarters: United States
Status: Public
Listed: IBM (opens google.com)NYSE
Valuation: public
Works in: QuantumSuperconductingComputing
Coverage: 3 programs · 13 updateslatest 4d agochecked 10 Oct
Customers: RIKEN (opens riken.jp)HSBC (opens hsbc.com)Basque Government (opens euskadi.eus)
Partners: AMD
People: Arvind Krishna (opens en.wikipedia.org)Chairman and CEOJay Gambetta (opens en.wikipedia.org)Director of IBM Research

Spec

NorthPole

Transistor count

22 billion^R (reported)

Core count

256^R (reported)

Operations per core per cycle

2,048 at 8-bit^R (reported)

Process node

12 nm^R (reported)

Die size

800 mm²^R (reported)

On-chip memory bandwidth

13 TB/s^R (reported)

Energy efficiency vs GPU

25x better (ResNet-50)^R (reported)