At Computex 2025, we met a startup doing one thing very totally different with AI chips. SemiQa is constructing an analog neural community processor designed for ultra-efficient sensor information processing, utilizing a customized reminiscence materials that would make dense, real-time edge AI much more accessible.
The issue with digital neural networks
As sensor information volumes develop, conventional processors wrestle to maintain up. Digital neural networks typically depend on complicated information conversion and devour extra energy than is good for edge functions.
SemiQa’s group believes the bottleneck isn’t in compute – it’s in information motion and conversion. In order that they’ve eradicated that step fully by staying within the analog area.
A customized materials with main price advantages
What makes SemiQa’s chip particular isn’t simply its structure — it’s the reminiscence itself. Their group developed a brand new analog reminiscence materials that’s cheaper to provide and simpler to scale than conventional memristors.
This offers them an edge in each efficiency and pricing, particularly in programs the place dense sensor inputs and energy budgets are main constraints.
The way it works: analog reminiscence + analog compute
The chip contains a matrix of analog reminiscence cells that retailer each incoming sensor information and neural community weights. It accepts a number of analog enter streams — like voice, vibration, or video — and processes them instantly with out changing to digital.
In observe, this implies no ADCs, no wasted energy, and no delay. The system can both classify inputs instantly or carry out quick, low-power preprocessing to information a digital neural community that follows.
Why you’d use this chip
SemiQa’s expertise is good for edge gadgets that have to course of steady sensor information with minimal energy drain. Instance functions embody:
- Sensible safety cameras with event-based wakeup
- Drones doing onboard imaginative and prescient or object recognition
- Battery-powered industrial sensors with always-on AI
And since the chip is programmable, it may be reconfigured within the discipline similar to a microcontroller — however optimized for dense AI duties.
A programmable MCU various for edge AI
Whereas it’s nonetheless underneath lively growth, this analog neural community chip may problem the dominance of microcontrollers in edge AI. SemiQa’s group confirmed a dwell demo at Computex with real-time analog reminiscence writes and reads, giving a glimpse into how shut that is to being productised.
For designers constructing next-gen edge programs, this might be the preprocessing engine that unlocks longer battery life, decrease latency, and extra clever sensor nodes.
Supply: ipXchange
