Power efficient methods for tackling machine learning workloads with next generation AI processors.
Real world applications involving AI workloads often comprise of perception and decision making. The AI system needs to percieve the environment in which it operates and make decisions based on the obtained information. Further the power constraints of edge devices and latency requirements of safety critical systems, necessitate the need for performing massive mathematical operations very efficiently. With this in mind at AlphaICs, we have developed the Real AI Processor (RAP) which is the fastest, most intelligent and power-efficient processor architecture for AI workloads. RAP’s architecture is based on “agent” as a basic compute element. Each agent consists of a group of “tensor compute units” thus enabling whooping performance for high dimensional computes. These agents operate in multi-agent environment for asynchronous processing of AI algorithms.
In this talk we shall provide a brief overview of the architecture of RAP and explain how AI workloads can be efficiently tackled using agent based compute. Different chipsets that AlphaICs is offering for both power efficient edge inferencing and efficient data center AI wokloads will be presented along with their performance numbers.