The Neural IP Portfolio is a comprehensive suite of components designed to support the development of Artificial Intelligence (AI) Systems on Chip (SoCs). This portfolio addresses the increasing performance requirements for AI systems, providing the industry’s highest performance and support for the latest, most complex neural network models. Here’s a brief introduction to the products:
- Datapath Elements: These are the fundamental building blocks for creating the data processing pathways in a neural network.
- Macro Cells: These are larger functional units in a neural network, typically composed of multiple datapath elements.
- Activation Functions: These are mathematical equations that determine the output of a neural processing unit.
- Multiply Accumulate Blocks (MAC): These are critical components in neural networks, performing the bulk of the computational work.
- Re-mapper with Low Power Cross-Bar: This is a specialized component designed to reconfigure data paths while minimizing power consumption.
- CNN Weight Memory: This is a memory unit specifically designed to store the weights of a Convolutional Neural Network (CNN).
- Pitch Match and GIGA Memory: These are specialized memory units designed for large-scale neural networks.
- Transpose Memory: This is a memory unit designed to support the transposition operations common in many neural network architectures.
These components are designed to address the demands of real-time compute with ultra-low power consumption for AI applications. The portfolio scales from 4K to 96K MACs and delivers, in a single instance, up to 250 tera operations per second (TOPS) at 1.3 GHz on 5nm processes in worst-case conditions, or up to 440 TOPS by using new sparsity features. This makes it suitable for a wide range of applications, including advanced driver assistance systems (ADAS), surveillance, digital TVs, cameras, and other emerging AI applications.