Uploaded November 2026 | Updated November 2026, 1 day ago
Blumind develops ultra-low-power analog AI for always-on sensing at the edge. Its BM110 chip processes microphone signals through analog feature extraction and neural processing, then sends a digital output to the system processor. Arthur Mason describes the architecture and its use in the Blumind EV kit. blumind.ai
The EV kit demonstrates voice activity detection with a microphone and LED indicator, along with keyword detection configured for Alexa. Blumind says the BM110’s analog compute uses about 25–30 microwatts for always-on detection, with power that stays roughly constant across workloads. The design avoids ADCs in its compute path and is intended to support long battery life in devices such as smart glasses and watches.
The discussion also covers energy and throughput figures: about four femtojoules per operation, and roughly five TOPS at around five milliwatts for the workloads described. The interview distinguishes those measured workloads from a normalized figure of 250 TOPS per watt. In a multi-stage system, the BM110 can handle low-power wake detection before higher-power processing runs on another chip or in the cloud.
Blumind’s next-generation architecture is in development. Mason describes plans for targeted speech extraction that detects a user’s voice, as well as scaling to larger models for vision processing, small language models, and transformer-based functions. The stated roadmap extends from hundreds of thousands of parameters toward 200 million, with longer-term targets of 3–10 billion parameters.
The BM110 is fabricated by GlobalFoundries in Dresden on a 22 nm process, according to the interview. At electronica 2026, Mason discusses the working silicon, its analog signal path, and how always-on sensing could connect low-power edge devices with more demanding local or cloud AI.
Blumind develops ultra-low-power analog AI for always-on sensing at the edge. Its BM110 chip processes microphone signals through analog feature extraction and neural processing, then sends a digital output to the system processor. Arthur Mason describes the architecture and its use in the Blumind EV kit. blumind.ai
The EV kit demonstrates voice activity detection with a microphone and LED indicator, along with keyword detection configured for Alexa. Blumind says the BM110’s analog compute uses about 25–30 microwatts for always-on detection, with power that stays roughly constant across workloads. The design avoids ADCs in its compute path and is intended to support long battery life in devices such as smart glasses and watches.
The discussion also covers energy and throughput figures: about four femtojoules per operation, and roughly five TOPS at around five milliwatts for the workloads described. The interview distinguishes those measured workloads from a normalized figure of 250 TOPS per watt. In a multi-stage system, the BM110 can handle low-power wake detection before higher-power processing runs on another chip or in the cloud.
Blumind’s next-generation architecture is in development. Mason describes plans for targeted speech extraction that detects a user’s voice, as well as scaling to larger models for vision processing, small language models, and transformer-based functions. The stated roadmap extends from hundreds of thousands of parameters toward 200 million, with longer-term targets of 3–10 billion parameters.
The BM110 is fabricated by GlobalFoundries in Dresden on a 22 nm process, according to the interview. At electronica 2026, Mason discusses the working silicon, its analog signal path, and how always-on sensing could connect low-power edge devices with more demanding local or cloud AI.










