Ceva introduced a new generation of NeuPro neural-processing designs on October 28, expanding its licensing portfolio into cloud and data-center infrastructure, according to its product announcement.
The company said the range uses a shared architecture and software development kit across four designs: Nano, Prime, Max and Ultra. The approach is intended to let semiconductor developers use common tools across different computing requirements.
Nano targets continuously operating sensors and basic perception tasks. Prime supports mainstream neural-network workloads, while Max addresses more demanding applications such as robotics, multimodal models and on-device assistants. Ultra targets infrastructure, including large-language-model inference and processing within memory and connectivity systems.
Ceva said the architecture offers up to twice the AI throughput per unit of chip area and a 50% improvement in energy efficiency compared with the first generation. Those figures are company-reported comparisons.
The company said selected early-access customers can use the architecture now, with general availability planned for the second quarter of 2027. It described semiconductor and system companies as evaluating and integrating the technology for future products; that does not establish widespread availability of finished customer devices today.
Sources: Ceva product announcement, distributed by PR Newswire.