The array becomes the chip
Multi-core architecture, with all the compute you need
99.9% accuracy on the most demanding AI workloads
Chips that work together
Embedded to edge to server
One platform for any model
Deploy from robot to rack
One chip becomes the whole stack.
Chip, card, server, rack, data center: one architecture at every scale.
Three ways to deploy.
-
-
-
Server
AI at data center scale: multi-AIPU cards that bring inference density to racks and supercomputers.
From Chip to Complete System
Choose how you deploy.
-
-
-
-
-
-
Embedded
One platform, multiple end-users
-
Industrial Manufacturing
Edge AI for machine vision and Industry 4.0 equipment, raising productivity in production lines and warehouses.
-
Retail
Real-time analytics across every customer touchpoint, so inventory control is easier and shrinkage reduces.
-
Security
Video analytics that turns a camera from a video source into a smart sensor for investigation and risk prevention.
-
Healthcare
High-performance AI at the edge keeps patient data safer while filling critical gaps in healthcare systems.
-
Smart Cities
Technology and data-driven solutions that make cities more efficient, sustainable and responsive to residents.
-
Agriculture
Better crop yields, optimized nutrients in the soil, and the right amount of water reaching the roots.
-
Space
Autonomous rovers navigating distant planets, and AI-powered satellites monitoring Earth and deep space.
-
Computer Vision
Vision moved to where the data is created: real-time processing, less data movement, scale into remote sites.
-
Robotics
Robots that enter non-permissive environments and make real-time decisions with no internet connection.
-
High Performance Compute
Digital In-Memory Compute scales up: the Titania chiplet serves the high performance compute market directly.
Three generations, one architecture
One architecture that spans embedded to rack-scale deployments
Metis®
Quad-core AIPU
- Typical application power 3.5-9 W
- Memory up to 16 GB per chip
- LLM capacity up to 8B parameters
- YOLOv8l 181 FPS
- ResNet-50 up to 3,200 FPS
Europa™
Double the cores, sixteen vector units, pre/post-processing on die.
- TDP 45 W per card
- Typical power 30-40 W
- Memory up to 64 GB per chip
- LLM capacity up to 32B parameters
- Qwen3-30B-A3B 53.5 tokens/s
Everything you need to build, faster.
One toolchain, every deployment: describe your pipeline in YAML and the Voyager SDK compiles, quantizes and runs it on any Axelera product line, with no retraining required.
Voyager SDK: the end-to-end software stack
Purpose-built for AI inference: a declarative YAML pipeline language, 100+ supported models through one toolchain, and API compatibility with industry standards. From idea to deployment quickly, with deep customization when you need it.
Axelera Wingman: the agentic AI platform
Purpose-built for Axelera hardware with deep, curated knowledge of the SDK and documentation: Axelera Wingman builds, debugs and optimizes inference pipelines from natural language, at 20 to 30% higher accuracy than general-purpose AI coding tools.
Describe the workflow in plain language; Voyager Wingman creates and deploys it on device.
The running pipeline, inspected on real Metis silicon without leaving the desk.
Partner Accelerator Network
From hardware to software to distribution, Axelera partners with leading companies to deliver complete AI solutions.
Latest News
.jpeg?width=560&height=310&name=Image%20(20).jpeg)
