AI Processing at 120mW? | Ambient Scientific

Modern models grow faster than batteries improve. Digital logic moves data, multiplies, adds, and repeats. Every step costs energy. Analog compute offers another path. It lets physics do part of the work.

Why MACs Eat Your Battery

AI inference is mostly multiply-accumulate. Digital units fetch from memory. They compute in the ALU. They write back. This is accurate and scalable. It is also power hungry at scale. Millions of parameters add up.

Analog Where It Matters

Ambient Scientific executes the MAC in analog. Voltages represent activations. Conductances represent weights. Currents sum by Kirchhoff’s law. Multiplication and addition become native circuit behaviour. Calibration and compensation manage variation and temperature. The digital side keeps control and flexibility.

Meet GPX10

GPX10 pairs a Cortex-M4 with custom AI cores. The analog MAC fabric handles the heavy lifting. The digital subsystem schedules work and I/O. Active inferencing targets roughly hundred-milliwatt power while sustaining high effective throughput. Always-on “subconscious mode” drops into microwatts. The device remains aware and ready to trigger.

This enables designs that do not sleep in the usual sense. They listen. They recognise. They wake the full system only when needed. That pattern suits IoT sensors, wearables, and appliances.

A Wearable That Actually Watches Out

In the demo, a pendant for elderly care detects falls. It fuses multiple sensors. It distinguishes a jump from a fall. It sends an alert when it matters. This is the promise of local AI on a tiny power budget.

Developer Experience

You code for a familiar MCU environment. You map MAC-heavy layers to the analog engine. You keep control flow and interfaces in digital. The result is low latency, reduced radio traffic, and better privacy. This is attractive for edge products where data must stay local.

Where It Fits

Consider battery devices that must be always aware. Asset trackers. Industrial monitors. Smart home sensors. Medical and consumer wearables. If your current design burns power in standby or offloads raw data, this architecture is worth a look.

Why It Matters Now

Digital scaling is slowing for energy per operation. Domain-specific hybrids are rising. GPX10 shows how analog plus digital can meet tight energy budgets without giving up programmability.

What To Do Next

Compare this approach with your current MCU + accelerator plan. Identify layers where MAC density dominates. Estimate duty cycles for an always-aware mode. Then prototype.