1. AI processor intellectual property aimed at TinyML.
- Subjects
SPEECH perception ,FEATURE extraction ,ARTIFICIAL intelligence ,INTELLECTUAL property ,MICROCONTROLLERS - Abstract
Ceva has developed an AI processor intellectual property (IP) that is specifically designed for System-on-Chips (SoCs) that need to run TinyML models. The IP comes in two versions: NPN32 with 32 int8 MACs and NPN64 with 64 int8 MACs. Both versions benefit from Ceva-NetSqueeze, which allows for direct processing of compressed model weights. The NPN32 is optimized for TinyML workloads targeting voice, audio, object detection, and anomaly detection, while the NPN64 provides 2x performance acceleration and is suitable for more complex on-device AI tasks such as object classification, face detection, speech recognition, and health monitoring. The IP includes sufficient hardware for neural network compute, feature extraction, running control code, and running DSP code without the need for an external microcontroller. Ceva's associated AI stack, NeuPro Studio, works with AI inference frameworks like TFLM and μTVM. [Extracted from the article]
- Published
- 2024