CEVA Joins Samsung's Foundry and Unveils NeuPro-M for Generative AI Inferencing

CEVA, Inc., a wireless connectivity and smart sensing technologies provider, announced its partnership with the Samsung Advanced Foundry Ecosystem (SAFE™). The collaboration aims to expedite chip design and reduce time-to-market for CEVA licensees utilizing Samsung's cutting-edge foundry processes.

The Samsung Foundry offers a comprehensive suite of advanced process technologies, including 28FD-SOI, 14/10/8/5/4nm FinFet, and 3nm GAA with EUV technology from 5nm. CEVA's IPs have already been integrated into Samsung's foundries across various sectors like 5G infrastructure, automotive, and consumer electronics. The partnership will further minimize supply chain risks by broadening the advanced manufacturing process options for CEVA's clientele.

The Samsung SAFE™ IP Partner Program is integral to the SAFE™ initiative, fostering a robust ecosystem between Samsung Foundry and IP partners. It offers a diverse IP portfolio tailored to performance-intensive applications.

Just last week, CEVA introduced its revamped NeuPro-M NPU family, addressing the increasing computational inferencing demands of Generative AI. The NeuPro-M NPU architecture has been reengineered to support transformer networks, Convolutional Neural Networks (CNNs), and other neural networks, facilitating optimized applications for both Generative and classic AI across various platforms.

NeuPro-M boasts an impressive 350 TOPS/Watt, catering to sectors from infrastructure to mobile with unparalleled cost and power efficiency.

Generative AI is set to boost the growth of Edge AI shipments, which are projected to surge from 2.4 billion units in 2023 to 6.5 billion by 2028. The challenges introduced by Generative AI in terms of performance, power, and memory on edge devices necessitate more potent and efficient AI inferencing techniques.

Reece Hayden of ABI Research highlighted the potential of CEVA's NeuPro-M NPU IP in deploying generative AI on-device within a tight power budget. The NeuPro-M NPU IP's scalability also makes it suitable for high-performance applications in network equipment and beyond.

The enhanced NeuPro-M architecture is versatile, supporting any future network layer and any activation or data flow. It offers up to 4X acceleration in performance, catering to diverse AI markets. The NeuPro-M family now includes the NPM11, NPM12, NPM14, and NPM18 NPUs, with a peak performance of 350 TOPS/Watt at a 3nm process node.

CEVA also unveiled a revamped development tool chain, based on its award-winning neural network AI compiler, CDNN. The software is compatible with popular open-source frameworks like TVM and ONNX.