Startups leverage analog computing and sparsity to run always-on edge AI workloads under one milliwatt

Startups and established IP designers are shifting edge AI workloads into the microwatt and low-milliwatt range to enable always-on sensing without draining batteries. Aspinity’s AML100 uses a reconfigurable analog modular processor (RAMP) architecture to analyze time-series data entirely in the analog domain, drawing just 30 microwatts to act as an "AI cache" that wakes up digital processors only when necessary. Similarly, Blumind utilizes single-transistor multiply-accumulate units to process continuous voltage patterns in analog, targeting under 100 microwatts for keyword spotting. In the digital domain, Femtosense’s AI-ADAM100 microcontroller leverages sparse memory and sparse compute to skip zero-value operations, achieving a combined 100x efficiency boost for voice processing. Efficient Computer takes a different approach with its Electron E1, a general-purpose CPU built on a spatial dataflow model that pins instructions to static tiles, claiming up to 100x the energy efficiency of traditional Arm cores. Meanwhile, Syntiant’s NDP250 runs deep neural networks up to six million parameters under 50 milliwatts, and Arm’s Ethos-U85 microNPU introduces native transformer support to the ULP edge. These diverse architectures demonstrate that ULP edge AI is moving away from traditional digital accelerators toward highly specialized, power-optimized silicon.

Aspinity's AML100 analog AI chip operates entirely in the analog domain, drawing only 30 microwatts of power to pre-filter sensor data. Blumind's BM110 analog processor performs keyword spotting under 100 microwatts using single-transistor multiply-accumulate units and capacitor-based storage.

Efficient Computer's Electron E1 general-purpose CPU utilizes a spatial dataflow model to achieve up to 100 times the energy efficiency of traditional architectures. Femtosense's AI-ADAM100 microcontroller implements sparse compute and sparse memory to deliver a 100-fold efficiency improvement for voice and audio processing.

Syntiant's NDP250 neural decision processor runs deep neural networks with up to six million parameters while consuming less than 50 milliwatts. Arm's Ethos-U85 microNPU brings native transformer support and up to 4,000 giga-operations per second of performance to ultra-low-power edge devices.