This research adapts large language models for Intel's Loihi 2 neuromorphic processor, creating a MatMul-free architecture that significantly reduces energy consumption while maintaining performance.

• Developed a 370M parameter MatMul-free model with no accuracy loss after quantization
• Achieved up to 3x energy efficiency improvement compared to traditional approaches
• Leverages event-driven computation and stateful processing capabilities of neuromorphic hardware
• Demonstrates practical hardware-aware quantization techniques for LLMs

This engineering breakthrough addresses a critical challenge in AI deployment by making powerful language models more energy-efficient and sustainable for real-world applications.

Neuromorphic Principles for Efficient Large Language Models on Intel Loihi 2