Ceva launches next-gen NeuPro NPUs with 2X AI throughput

A new tier of AI hardware

Ceva has unveiled its next-generation NeuPro Neural Processing Unit (NPU) portfolio to handle increasingly complex AI workloads. These NPUs are specialized chips designed to accelerate neural networks, which are the mathematical foundations of modern artificial intelligence. The new lineup aims to support multimodal, generative, and agentic AI applications across a wide variety of hardware environments.

The company reports that these new chips deliver up to 2X higher AI throughput per mm² compared to the first generation. This means the hardware can process twice as much data within the same physical space on a silicon chip. Additionally, the new portfolio offers a 50% improvement in energy efficiency, which is critical for battery-powered devices.

To achieve these gains, Ceva utilized advanced AI numerics and mixed precision, specifically including FP4 and FP8 formats. These formats allow the chips to perform calculations with less precision than standard methods, which significantly reduces the amount of data that needs to move between memory and the processor. This optimization helps the hardware manage the heavy demands of transformer-based models while staying within power limits.

Scaling from sensors to clouds

The new NeuPro portfolio is organized into four distinct tiers to serve different market needs. NeuPro Nano is designed for always-on sensing and basic perception in devices with very tight space and power constraints. NeuPro Prime targets mainstream workloads involving Convolutional Neural Networks (CNNs) and transformer models, which are common in everyday AI tasks.

For more demanding needs, NeuPro Max scales to advanced edge AI applications. This includes high-resolution perception, on-device assistants, robotics, and Physical AI systems. These systems require the ability to process complex information in real-time, such as a robot interacting with its physical surroundings or a high-end wearable processing multiple data streams simultaneously.

At the top of the lineup is NeuPro Ultra, which marks Ceva's expansion into AI infrastructure. This chip is built for large language model inference in cloud and data center environments. It also supports intelligent memory and connectivity silicon, which helps manage data movement within a system. By putting neural processing into these subsystems, Ceva aims to reduce bottlenecks where data is stored rather than just where it is processed.

Unified software and architecture

One of the primary goals of the new portfolio is the use of a common core architecture and a unified software development kit (SDK). This means that semiconductor designers can use the same tools and software investments across all four NPU types. They can choose the specific engine that fits a particular part of their system or combine multiple engines within a single System on a Chip (SoC).

This unified approach helps companies reduce the complexity of integrating different AI components into their products. Instead of learning a new software stack for every different chip, developers can move between cores as their requirements change. This consistency helps shorten the development cycle for successive generations of products, allowing companies to scale from simple sensors to complex edge systems more easily.

Ceva's strategy focuses on the shift toward hybrid AI, where perception happens near the sensor and complex reasoning is distributed. Because different parts of this system have different constraints—measured in microwatts for sensors and megawatts for data centers—the unified architecture provides a consistent way to develop for both. CEO Amir Panush said a common stack lets customers build once and deploy across their roadmap.

Technical optimizations for models

The next-generation NeuPro architecture includes several specific features to handle modern AI model architectures. It supports Mixture-of-Experts and linear attention, which are techniques used to make large models more efficient. The architecture is also designed with enough flexibility to accommodate new model types that have not yet been invented, ensuring the hardware remains useful over a longer product lifecycle.

To improve performance, the chips utilize optimized tensor reshaping to maximize multiply-accumulate (MAC) utilization. MACs are the basic operations performed by neural networks; by optimizing how these tensors are shaped, the hardware can ensure that more of its computing power is actually being used for useful work. This leads to better inference performance without needing to increase the physical size of the chip.

Furthermore, the portfolio supports pre-quantized models and advanced compression, quantization, and pruning. These techniques reduce the amount of computation and data movement required for a model to run. By supporting pre-quantized models, customers can deploy published checkpoints immediately without having to perform a separate and often difficult calibration step. These features collectively improve the effective performance per watt and per unit area for developers.

Market outlook and availability

The expansion into AI infrastructure is a strategic move prompted by the growing demands of cloud AI inference. ABI Research forecasts that cloud AI inference capacity will grow at a 42% CAGR through 2035. As these workloads scale, the movement of data between memory, compute, and networks becomes a primary bottleneck for system performance and power efficiency.

NeuPro Ultra is specifically designed to address these infrastructure challenges by supporting multi-user operations and high memory bandwidth. It can be integrated into networking subsystems to apply neural processing to the movement of data itself. This reduces unnecessary transfers and relieves bottlenecks closer to where the data is stored, which is essential for running large-scale models for many users at once.

Ceva is already engaging with more than 15 customers who licensed the first generation of NeuPro to help shape this next generation. While the architecture is currently available to select early-access customers, Ceva plans for general availability in the second quarter of 2027. The company will discuss these new products and its strategy for Physical AI at upcoming technology symposiums in Taiwan and China.