The World’s Most Energy Efficient AI Processor
Revolutionary AI processing technology for hyperscalers, inference providers, and major AI users.
Ultra Power Efficient
Breakthrough energy efficiency that reduces operational costs by up to 70% compared to traditional processors.
Scalable Architecture
Modular design that scales seamlessly from edge devices to massive data center deployments.
AI-Optimized
Purpose-built for AI workloads with specialized tensor processing units and optimised memory architecture.
Cost of LLMs
Today’s AI revolution faces a critical challenge. While large language models have shown remarkable capabilities in understanding and generating human language, their deployment comes at an enormous cost.
Large language models like ChatGPT require huge data centers full of expensive GPUs and consuming enormous amounts of electricity.
The Data Centre Dilemma
Current AI models require massive data centers, with facilities like OpenAI’s planned infrastructure spanning up to 10 square miles and consuming power equivalent to five nuclear reactors.
Not only do they consume vast amounts of electricity but also precious freshwater used for cooling.
Minimal Source Code Changes
Individual large data centers can consume electricity equivalent to cities of 1.8 million people , creating infrastructure requirements that only the largest technology companies (and by extension, the most technologically advanced nations) can afford. This creates not just a corporate divide, but a sovereign technology gap that threatens national competitiveness in the AI era.
The Solution: Smarter Not Bigger
The Power of Mathematics
Instead of building ever-larger computers, we’ve developed a fundamentally new approach to AI computation. Our innovation centers on a simple insight: AI models don’t need perfect precision for every calculation . By using a simplified (ternary) number system for certain operations, we achieve remarkable efficiency gains
Less Memory
This ternary approach means AI models require significantly less memory, allowing the same hardware to run much larger and more capable models—or enabling smaller, cheaper devices to run models that previously required expensive server hardware. Complex multiplication operations are replaced with simple addition and subtraction, dramatically simplifying the required hardware.
Optimised for Purpose
Unlike GPUs which are designed for general-purpose computing and carry significant overhead, our ASIC design is optimized specifically for ternary operations. This eliminates the wasted silicon area and power consumption that GPUs require for floating-point units, complex instruction decoders, and other general-purpose features.
Target Markets
Hyperscalers
Power massive cloud AI services with unprecedented efficiency and performance.
AI Inference Providers
Deliver faster, more cost-effective AI inference at any scale.
Enterprise AI Users
Transform business operations with efficient on-premise AI processing.
The Challenge: AIs Growing Resource Crisis
Take the deck with you.
Seen us pitch? The slides and the live demo are here whenever you need them.
The pitch deck
The slides from our investor presentation, as a PDF for later reference.
Watch it run
The same model on the same FPGA, recorded live. BF16 on the left, ternary on the right.