Blumind has released the BM110 Evaluation Kit, giving developers a hardware and software platform to test always-on audio and sensor inference using its AMPL analog computing technology. The kit combines BM110 silicon, preloaded keyword-spotting and voice-activity-detection models, a software development kit and onboard power measurement infrastructure. The launch targets a growing challenge in edge AI: enabling continuous intelligence in battery-powered devices without exhausting their limited energy budgets.
Always-on AI is becoming an important requirement for wearable devices, smart home products and other connected systems that need to recognize activity or respond to voice commands without waiting for a user to activate them. But continuously running neural networks can place significant demands on battery life, particularly when devices must monitor microphones and sensors throughout the day.
Blumind is targeting this challenge with its BM110 Evaluation Kit, a development platform designed to let semiconductor customers test always-on inference using the company’s AMPL analog computing technology.
The kit includes BM110 silicon configured as an always-on audio reference board, along with a software development kit, preloaded reference models and infrastructure for measuring current consumption directly on the hardware. Blumind says the platform allows engineering teams to evaluate power consumption and inference performance under real operating conditions rather than relying only on demonstrations or theoretical specifications.
Analog Computing Targets the Energy Cost of AI Inference
Many conventional neural network accelerators use digital processing architectures in which model weights are stored in memory and repeatedly transferred to processing units during computation.
Moving data between memory and processing elements consumes energy. For devices that perform occasional AI tasks, this overhead may be manageable. In always-on applications, however, continuous data movement and processing can become a significant constraint.
Blumind’s AMPL technology takes a different approach by storing neural network weights directly in the BM110 analog compute fabric. The company says this reduces the need to repeatedly move model parameters between memory and a digital multiply-accumulate processing array.
The BM110 neural network compute core is designed to operate continuously at less than three microwatts, according to Blumind. The company reports total always-on system power of approximately 50 microwatts, including the microphone.
These figures are central to the chip’s positioning. Wearables such as smart glasses, smartwatches and rings have limited battery capacity and little room for additional thermal or power overhead. A dedicated low-power inference component could allow these devices to monitor audio continuously while preserving energy for their other functions.
However, the reported figures are vendor specifications. Actual power consumption will depend on the model, microphone configuration, operating conditions and application workload. Independent comparisons against competing digital inference architectures would help establish the scale of the advantage.
BM110 Evaluation Kit Makes Power Measurement Part of Development
The evaluation kit is designed to give developers direct access to the hardware and software needed to test the BM110.
Its reference board includes the BM110 chip, an analog microphone and an integrated signal-processing front end. Battery management circuitry and a series sense resistor provide onboard current measurement, allowing developers to observe system power consumption in real time.
This measurement capability is important because always-on AI performance cannot be evaluated solely through inference accuracy or processing speed. Energy consumption during continuous operation is a defining metric for battery-powered devices.
The kit also includes preloaded models for keyword spotting and voice activity detection (VAD). Keyword spotting enables a device to recognize specific trigger words, while VAD identifies whether speech is present in an audio stream.
These are common building blocks for voice-enabled products. They can support functions such as wake-word detection, voice activation and low-power audio monitoring before a larger processing system is engaged.
Blumind’s software development kit includes a hardware abstraction layer intended to separate application code from the host platform and support integration with different microcontrollers.
The company also provides a model import workflow for networks trained using standard cloud and server frameworks. This is intended to reduce the amount of work required to move trained models into the BM110 environment.
For engineering teams, compatibility with existing model development workflows could be important. A specialized inference chip has limited value if developers must rebuild their machine-learning pipelines or adopt an entirely separate training environment to use it.
Analog AI Raises Questions About Accuracy and Long-Term Stability
Low power is only one part of the evaluation process. Developers also need to understand model accuracy, programming precision, temperature behavior and the stability of stored neural network weights over a product’s lifetime.
Blumind says the BM110 kit is designed to help customers investigate these characteristics using real hardware and live inference workloads.
Qualified customers can also request the company’s AMPL Technology Validation Report. According to Blumind, the report covers weight storage and programming accuracy across process, voltage and temperature conditions, as well as modeled weight drift over product lifetime and end-to-end system power for representative always-on workloads.
These measurements address important engineering questions associated with analog computing. Analog circuits can be affected by manufacturing variation, operating conditions and changes in component characteristics. Maintaining predictable model behavior across these variables is essential for commercial deployment.
The availability of validation material could help developers assess whether the architecture meets their product requirements. However, the announcement does not disclose detailed accuracy figures, model sizes, supported neural network architectures or comparative benchmark results against established low-power digital accelerators.
Those details will be important when customers evaluate the BM110 for specific applications.
Physical AI Expands Demand for Continuous Edge Intelligence
Blumind’s launch reflects a broader shift toward edge AI systems that remain active rather than responding only to occasional commands.
Physical AI applications increasingly depend on continuous sensor input to interpret surrounding environments and respond to events. Audio sensing is one example, but similar energy constraints apply to motion detection, environmental monitoring and other sensor-driven workloads.
For wearable devices, always-on inference could support voice interfaces and context-aware interactions without continuously activating a more power-intensive application processor. In connected industrial equipment, low-power sensor inference could enable persistent monitoring while limiting energy consumption.
The BM110 Evaluation Kit gives developers a way to test whether analog inference can deliver the power efficiency required for these applications.
The kit is available now to qualified customers, with Blumind providing its datasheet, technology validation documentation and direct applications engineering support through model bring-up. The company plans to showcase the platform at electronica 2026 in Munich.
The broader commercial question is whether analog compute can combine its potential energy advantages with the accuracy, stability and software compatibility required by production devices. For Blumind, the evaluation kit is an opportunity to move that discussion from architecture claims to measurable engineering results.
Market Landscape
The edge AI semiconductor market is increasingly focused on reducing the energy and processing requirements of continuous inference. Wearables, smart home devices and sensor-driven industrial systems need to analyze incoming data without relying on power-intensive processors for every operation.
Digital neural network accelerators remain an established approach, but analog computing architectures offer an alternative by reducing data movement and performing computation closer to where model weights are stored.
Blumind’s BM110 targets this opportunity through its AMPL technology and a development kit built around direct power measurement. Its market prospects will depend on inference accuracy, model compatibility, long-term stability, integration with microcontrollers and independently validated power performance.
Top Insights
- Analog neural computing: BM110 stores neural network weights in its analog compute fabric to reduce energy-intensive data movement.
- Low-power claims: Blumind reports less than three microwatts for the neural compute core and approximately 50 microwatts for the complete always-on system, including the microphone.
- Developer evaluation: The kit includes BM110 hardware, an SDK, model import tools and preloaded audio inference models.
- Direct measurement: Onboard current-sensing infrastructure lets engineering teams measure power consumption during active inference.
- Commercial validation: The AMPL validation report addresses programming accuracy, weight drift and performance across operating conditions.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
