Executing spiking neural networks (SNNs), BrainChip’s licensable Akida NPU can be more power and area efficient than NPUs for convolutional neural networks (CNNs). However, converting a CNN to an SNN lessens Akida’s advantage and entails additional developer complexity. BrainChip, therefore, is leaning into its advantage, offering a new scaled-down Pico configuration that can operate without a supporting control CPU and focusing on neural networks that natively map to its hardware.
The licensable design (IP) should allow always-on battery-powered electronics to run many times longer than other NPUs, particularly if they depend on a CPU or DSP. BrainChip has collaborated with Edge Impulse to help customers train and deploy models, such as for biometric analysis in wearable medical and health devices. The IP supplier also targets audio processing such as keyword spotting and noise reduction for voice-controlled gadgets, earbuds, hearing aids, and similar designs.
If Akida Pico proves successful, we expect the company to pair different Akida configurations with models for other functions, such as retrieval-augmented generation (RAG) scaled down for edge applications or image recognition. An SNN (or TENN, BrainChip’s term that deemphasizes SNNs’ biomimicry) could have further advantages for the latter because image detection is intrinsic to SNN’s event-driven nature and latency can be reduced compared with a CNN, increasing the distance at which an object is recognized.
Turning away from competing with other NPU suppliers on their terms to focus on its technology’s better power efficiency when running optimized models, BrainChip seeks to stand out in a market crowded with numerous startups and a few big companies.