Idea
Sparse MoE model delivering dense-level accuracy with reduced compute and storage for efficient edge device AI.
Research Paper
Core Innovation
This paper introduces DECO, a sparse MoE architecture using ReLU-based routing with learnable expert scaling and a novel NormSiLU activation. It achieves dense Transformer-level performance with only 20% expert activation and simplifies MoE design by using non-gated MLP experts, improving efficiency and stability.
Why It Matters
End-side devices require AI models that balance high performance with low computational and storage demands. DECO addresses this by reducing memory bottlenecks and computation while maintaining accuracy, enabling advanced AI capabilities on resource-constrained hardware. This improves user experience and broadens AI adoption in mobile, IoT, and embedded systems.
Market Size (TAM)
$20–50B TAM for edge AI hardware and software; $2–10B SAM from mobile and IoT device manufacturers. Driven by demand for efficient AI inference and growing edge computing adoption.
Potential Customers & Pain Points
- Mobile device manufacturers – Need efficient AI models with low storage and compute
- IoT solution providers – Require high-performance models on limited hardware
- Edge AI developers – Face memory and speed bottlenecks in deploying large models
Business Model
Licensing the DECO architecture and specialized acceleration kernel to device manufacturers and AI platform providers; offering consulting and integration services for edge AI deployment.
Competitive Landscape
- Google Edge TPU
- NVIDIA Jetson
- Qualcomm AI Engine
- OpenAI MoE models
Implementation Challenges
- Integration complexity with existing edge AI pipelines
- Hardware compatibility and optimization challenges
- Competition from established efficient model architectures
Validation Strategy
- Benchmark DECO on diverse edge devices against dense and existing MoE models
- Partner with hardware vendors for real-world deployment pilots
- Collect performance
- power consumption
- and user experience data to refine model and kernel
Research Paper Overview
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
Summary
DECO is a sparse Mixture-of-Experts architecture that matches dense Transformer performance while reducing computation and storage, enabling efficient deployment on end-side devices. It activates only 20% of experts and achieves a 3x speedup on real hardware, addressing memory bottlenecks and computational cost without sacrificing accuracy.