CEVA Launches Enhanced NeuPro M NPU IP Series, Vigorously Promoting Generative AI

NeuPro-M provides industry-leading 350 TOPS/Watt performance, with exceptional cost and energy efficiency as its infrastructure

Powerful Generative AI brought by industrial, automotive, PC, consumer products, and mobile markets

CEVA, Inc. (NASDAQ:. The NeuPro-M NPU architecture and tools have been redesigned to support not only CNN and other neural networks, but also Transformer networks and future machine learning inference models, making it possible to use them in communication gateways, optical connectivity networks, cars, laptops and tablets, AR/VR headworn devices, smartphones, and any other cloud or edge use cases, Seamless development and highly optimized application based on NeuPro-M NPU, fully utilizing the functions of generative artificial intelligence and traditional artificial intelligence.

Ran Snir, Vice President and General Manager of the Visual Business Unit at CEVA, commented: Transformer based networks require a significant increase in computing and memory resources to drive generative artificial intelligence, requiring new methods and optimized processing architectures to meet this growing demand for computing and memory. Our NeuPro-M NPU IP is specially designed to efficiently and cost-effectively handle current and future traditional artificial intelligence and generative artificial intelligence workloads. This IP is scalable and can handle tasks from the edge to the cloud End use cases, and can support new inference models in the future. We have achieved a performance leap using this architecture, bringing a beautiful prospect for generative artificial intelligence for various use cases, covering everything from cost sensitive edge devices to efficient cloud computing, and all applications in between

Market research firm ABI Research predicts that the shipment volume of edge artificial intelligence will increase from 2.4 billion units in 2023 to 6.5 billion units in 2028, with a compound annual growth rate (CAGR) of 22.4% *. Generative artificial intelligence will be an important force supporting growth, and increasingly complex intelligent edge applications are driving the demand for stronger and more efficient artificial intelligence inference technologies. Specifically, although the Large Language Models (LLMs) and visual and audio transformers used in generative artificial intelligence can change products and industries, running on edge devices will pose new challenges to performance, power consumption, cost, latency, and memory.

Reece Hayden, Senior Analyst at ABI Research, stated: Nowadays, the hardware market for generative artificial intelligence is highly concentrated, with a few manufacturers occupying a dominant position. In order to fulfill technological commitments, clear paths are needed in both the cloud and edge domains to achieve lower power consumption and lower cost inference processing, which will be achieved through smaller model sizes and more efficient hardware. CEVA's NeuPro-M NPU IP is a particularly attractive solution that can be deployed on devices with excellent power budgets Generative artificial intelligence, while also having scalability, enables NeuPro-M to meet more performance intensive use cases in network devices and other devices.

Benefiting from the continuous development of reasoning and modeling technology, people can utilize the new features of smaller specific fields such as LLM, visual transformer, and other generative artificial intelligence models at the device level, which will correspondingly change the applications in infrastructure, industry, automobiles, PCs, consumer products, and mobile markets. Most importantly, due to the integration of Vector Processing Units (VPUs), the enhanced NeuPro-M architecture has high versatility and future scalability, which can support any future network layer. In addition, this architecture also supports any activation function and any data stream, with true sparsity of data and weights, which can improve processing performance up to four times, allowing customers to use a single NPU series to meet the needs of multiple applications and markets.In order to achieve greater scalability required for a diversified artificial intelligence market, NeuPro-M has added new NPM12 and NPM14 NPU cores, each with two and four NeuPro-M engines, which can easily provide higher performance artificial intelligence workloads. The enhanced NeuPro-M series currently includes four NPUs, namely NPM11, NPM12, NPM14, and NPM18. NeuPro-M has versatility, excellent performance, and energy efficiency, making it a leading NPU IP product in the industry today. It can achieve 350 TOPS/Watt peak performance on 3nm process nodes, and can process over 1.5 million tokens per watt in transformer based LLM inference.

CEVA provides an improved comprehensive development toolchain for the enhanced NeuPro-M architecture, based on CEVA's award-winning neural network AI compiler CDNN. It has architecture aware capabilities and can fully utilize the NeuPro-M parallel processing engine to maximize customer AI application performance. The CDNN software includes a memory manager for reducing memory bandwidth and optimizing load balancing algorithms, and is compatible with common open source frameworks including TVM and ONNX.

Supply

NPM11 NPU IP is now fully available for customer deployment. CEVA provides NPM12, NPM14, and NPM18 to major customers. For more information, please visit the company website https://www.ceva-dsp.com/product/ceva-neupro-m/ .

*Source: ABI Research, AN-5774, QTR 1 2023

About CEVA Company

CEVA is a leading licensor of wireless connectivity and intelligent sensing technology as well as co creation solutions, aiming to create a smarter, safer, and more connected world. We provide digital signal processors, artificial intelligence processors, wireless platforms, encryption cores, and supporting software for sensor fusion, image enhancement, computer vision, voice input, and artificial intelligence applications. These technologies, along with our Intrinsix IP integration services, are provided to customers to help them solve complex and time critical integrated circuit design projects. Many of the world's leading semiconductor manufacturers, system companies, and OEMs utilize our technology and chip design skills to develop energy-efficient, intelligent, and secure interconnected devices for various end markets such as mobile, consumer, automotive, robotics, industrial, aerospace, defense, and the Internet of Things.

CEVA's DSP based solutions include 5G baseband processing platforms in mobile, IoT, and infrastructure; Advanced imaging technology and computer vision for camera equipment; Audio/voice/voice applications suitable for multiple IoT markets and ultra low power always on/sensing applications. For motion sensing solutions, our Hillcrest Labs sensor processing technology provides a wide range of sensor fusion software and inertial measurement unit ("IMU") solutions for markets such as headphones, wearable devices, AR/VR, PCs, robots, remote controls, and the Internet of Things. In terms of wireless Internet of Things, our Bluetooth connectivity (low power and dual mode), Wi Fi 4/5/6 (802.11n/ac/ax), ultra wideband (UWB), NB IoT, and GNSS platforms are widely authorized connection platforms in the industry.