Neuromorphic Processor IP (Second Generation)

Self-contained neural processor

• Scalable fabric of 1-128 nodes
• Each neural node supports 128 MACs
• 8,4,1, bit arithmetic precision
• Programmable activation functions
• Skip connections
• Configurable 50-130K embedded local SRAM
• DMA for all memory and model operations
• Multi-layer execution without host CPU
• AXI bus interface

 Efficient algorithmic hybrid mesh

• Performs as Temporal Neural Processor (TNP), spatial Convolutional Neural Processor (CNP) and Fully-connected Neural Processor (FNP)
• Integrates CNNs, Spatio-Temporal, TENNs Buffer networks

Akida efficiently accelerates…

• Image and audio classification
• Object detection
• Scene segmentation
• Gesture and face recognition
• State-of-the-art algorithms in sequence prediction
  • Video object detection
  • Human action recognition
  • Raw-audio classification
  • Vital signs prediction

Notable features:

• Supports 8-,4-,and1-bit weights and activations
• Supports multiple layers simultaneously
• Supports long-range skip connections in hardware

Software development and deployment:

• Akida leverages standard frameworks and development platforms such as TensorFlow/Keras,  Pytorch/ONNX and Edge Impulse
• Akida is model-, network-, and OS-agnostic
• BrainChip MetaTF supports model development and optimization for Akida hardware
• Akida model zoo offers a set of pre-built Akida-compatible models, pre-trained weights and training scripts
• Akida TENNs models are offered for evaluation

• Accelerates today’s networks: CNNs, DNNs and more, directly in hardware with minimal CPU intervention
• Event-based processing: Computes only when necessary; substantially reduces number of operations executed and energy consumed
• At-memory compute: Significantly reduces memory movement; uses cost effective, scalable, standard RAMs
• Exceptional spatio-temporal capability: Patented Temporal Event-based Neural Nets (TENNs) revolutionize time-series data applications
• Event-based communication: Sends data between NPUs through integrated mesh without any CPU intervention; offloads system
• Intelligent runtime: Runtime manages all operation of neural processor, transparent to the user, accessible through a simple API

• Fully synthesizable RTL.
• IP deliverables package with standard EDA tools.
• Complete test bench with simulation results.
• RTL synthesis scripts and timing constraints. Customized IP packaged targeted for your application.
• Run time software C++ library.
• Processor and OS agnostic.