AlphaIC’s RAPTM architecture consists of multiple RAPTM agents and each agent has one scalar processor, multiple tensor processors and dedicated on-chip memory. The scalar processor interacts with the host processor via AXI interfaces. Scalar processors fetch the instructions, decode it and offload the AI specific operations to the Tensor Processors. Tensor processors contain specialized compute logic for matrix multiply, activations and pooling. The architecture tightly couples agent and memory thus resulting in high performance AI operations
Based on a novel hardware that utilizes small tensor units and a novel instruction set architecture that minimizes overhead and increases efficiency, AlphaICs solution provides best-in-class performance.
AlphaICs software stack (AlphaRT) provides seamless environment for deploying neural networks onto the RAPTM. AlphaRT supports TensorFlow, and we plan to add support for other AI frameworks in the future.