In this episode of the AI Hardware Podcast, hosts Ian Cutras and Sel Foxton delve into the world of neuromorphic chips, exploring various companies and their approaches to brain-inspired computing. They begin with BrainChip, a long-standing player in the neuromorphic space known for its event-driven digital architecture aimed at ultra-low power sensor edge applications. BrainChip is evolving from traditional spiking models toward more modern state space models and integrating with mainstream AI software frameworks like TensorFlow and PyTorch, signaling a shift toward handling more complex tasks such as language processing.
Next, the discussion turns to Intel’s neuromorphic efforts with their Loihi chips, particularly Loihi 2. Unlike BrainChip’s sensor-edge focus, Intel’s neuromorphic chips are designed for large-scale systems, potentially forming neuromorphic supercomputers by networking thousands of small, low-power chips. Despite significant research progress and government funding, Intel’s neuromorphic projects remain largely academic with no clear path to commercialization, and the software ecosystem, including the Lava framework, is still maturing.
The hosts then highlight startups like SinSense and Spec, which focus on asynchronous digital neuromorphic chips for real-time, low-power sensor applications such as gesture recognition and occupancy detection. These companies combine event-based sensors with neuromorphic processors to create commercial products, primarily targeting edge AI markets. While these chips offer promising always-on capabilities, challenges remain in power consumption due to analog-to-digital conversion and the early stage of their software and developer ecosystems.
Another notable project discussed is Spica, originating from the University of Manchester and led by Steve Furber, a pioneer in ARM architecture. Spica’s neuromorphic chips leverage ARM cores optimized for event-driven workloads and emphasize scalable networking to build large neuromorphic systems. Recently, a German startup has licensed this technology to develop neuromorphic supercomputers aimed at applications like drug discovery, marking a significant step toward practical, large-scale neuromorphic deployments in the cloud.
The episode concludes with reflections on the definition of neuromorphic computing, emphasizing the importance of spiking neural models to truly capture brain-like behavior. The hosts suggest that companies claiming to be neuromorphic should incorporate spiking architectures to maintain credibility. Overall, the discussion underscores the diversity of neuromorphic approaches—from tiny sensor-edge chips to large-scale supercomputers—and highlights ongoing challenges in commercialization, software development, and defining the field’s boundaries.
