The Quadric Chimera General Purpose Neural Processing Unit (GPNPU) delivers unparalleled performance for AI workloads, characterized by its ability to handle diverse and complex tasks without requiring separate processors for different operations. Designed to unify AI inference and traditional computing processes, the GPNPU supports matrix, vector, and scalar tasks within a single, cohesive execution pipeline. This design not only simplifies the integration of AI capabilities into system-on-chip (SoC) architectures but also significantly boosts developer productivity by allowing them to focus on optimizing rather than partitioning code.
The Chimera GPNPU is highly scalable, supporting a wide range of operations across all market segments, including automotive applications with its ASIL-ready versions. With a performance range from 1 to 864 TOPS, it excels in running the latest AI models, such as vision transformers and large language models, alongside classic network backbones. This flexibility ensures that devices powered by Chimera GPNPU can adapt to advancing AI trends, making them suitable for applications that require both immediate performance and long-term capability.
A key feature of the Chimera GPNPU is its fully programmable nature, making it a future-proof solution for deploying cutting-edge AI models. Unlike traditional NPUs that rely on hardwired operations, the Chimera GPNPU uses a software-driven approach with its source RTL form, making it a versatile option for inference in mobile, automotive, and edge computing applications. This programmability allows for easy updating and adaptation to new AI model operators, maximizing the lifespan and relevance of chips that utilize this technology.
