AMLD Intelligence Summit 2026 – Pushing Biosignal FMs to the Edge
Presenting the FM edge-deployment poster at AMLD 2026Abstract
Presented our poster on Pushing Biosignal Foundation Models to Ultra-low-power Edge Devices, a unified full-stack framework for deploying EEG and EMG foundation models on the GreenWaves Technologies GAP9 ultra-low-power RISC-V microcontroller. We showcased the deployment path for four architectures: CEReBrO—a compact encoder using alternating attention that achieves the fastest inference at 146 ms and just 8 mJ per window; FEMBA—a bidirectional Mamba FM quantized down to 2-bit weights using QAT; TinyMyo—a 3.6 M-parameter Vision Transformer for gestures, kinematics, and silent speech running at ~36 mW always-on; and LUNA—our topology-agnostic EEG foundation model, now deployed on-device with channel unification via RoPE, FFT, and learned queries. The pipeline spans PyTorch training through Brevitas quantization-aware training, INT8 export, custom parallel C kernel generation, and hierarchical L3→L2→L1 weight streaming with double-buffered DMA on GAP9. Quantization compresses models up to 9× (FP32→INT8) with less than 2% accuracy drop, and the neural accelerator (NE16) boosts throughput to over 10 MACs/cycle—enabling the first real-time deployment of state-of-the-art biosignal foundation models on an ultra-low-power edge device.
Poster highlights
- Four foundation models, one pipeline: CEReBrO, FEMBA, LUNA, and TinyMyo—spanning EEG abnormality detection, seizure monitoring, EMG gesture recognition, and silent speech—all deployed through a unified PyTorch → Brevitas QAT → GAP9 pipeline.
- First real-time FM on GAP9: CEReBrO achieves 146 ms latency and 8 mJ per inference; FEMBA is quantized down to 2-bit weights; TinyMyo runs at ~36 mW always-on; and LUNA brings topology-agnostic channel unification to the edge.
- 9× compression, minimal accuracy loss: INT8 quantization across all models yields less than 2% accuracy drop on TUAB and CHB-MIT, with the NE16 neural accelerator pushing throughput beyond 10 MACs/cycle.
- Edge-ready wearables: BioGap integration projects up to 2 days of battery life, making continuous health monitoring with foundation-scale intelligence a practical reality.
Photo gallery
Poster & resources
Related work
Thorir Mar Ingolfsson
Postdoctoral Researcher
Making AI run on microwatts: I build foundation models for EEG, ECG, and other biosignals, and compress them until they run in real time on wearable devices, not in the cloud.