Akida 2 Technology Platform is a scalable Neural Processor Unit (NPU) for accelerating edge AI applications and use cases for embedded platforms. The Akida 2 features a unique neuromorphic processing paradigm which can significantly reduce the computational complexity by exploiting sparsity in the data sets and neural model. This results in Akida to only compute those portions of the data which can contribute to a result with a spiking or event based neural network. This can convert general purpose CNN networks to efficient event-based neural networks that can achieve orders of magnitude improvements on power efficiency. This key capability enables an order of magnitude reduction in the inference power needed. Neuromorphic compute has traditionally been an analog function, BrainChip’s key Akida innovation has been designed as a completely digital process, enabling its integration into standard digital System-on-Chip designs.
The Akida 2 Platform is scalable with a mesh network of NPU computing nodes which distributes memory and computations across the array. This enables efficient at-memory computations of each layer within that node and the mesh fabric can provide scalable compute resources to expand the performance and size of the network being processed. The combination of efficient event-processing, at-memory compute and scalable compute fabric provides a solution for consumer and wearable product developers looking for the ultimate in power efficient edge AI processing. Akida 2’s smallest implementation is the Akida pico, a sub-1 mm2, sub-1 mWatt core that can support the tiniest cost and power budgets.
BrainChip’s Akida extends edge AI processing to the lowest power levels achievable on a digital process to drive further on-device penetration that is low power and low cost, enabling integration into smart consumer and home products.