Ceva has introduced its next-generation NeuPro NPU portfolio, expanding the family beyond edge devices to include “AI infrastructure” use cases aimed at cloud and data center inference and silicon that sits closer to memory, connectivity, and data processing. Ceva is pitching the architecture as a single platform that can scale from always-on sensing up through larger inference deployments, with a unified software development kit across the lineup.
Ceva lists four NPUs in the portfolio: NeuPro Nano, NeuPro Prime, NeuPro Max, and NeuPro Ultra. Across the next-generation architecture, Ceva states up to 2X higher AI throughput per mm² and a 50% improvement in energy efficiency versus its first-generation NeuPro. The company also calls out mixed-precision support including FP4 and FP8, higher memory bandwidth, and memory-management features including dynamic shapes and paged KV cache, features that map directly to memory-bound transformer inference behavior.
On the model side, Ceva says the architecture supports Mixture-of-Experts and linear attention, along with “support for pre-quantized models” so customers can deploy published checkpoints without a separate quantization and calibration step. It also lists higher multiply-accumulate utilization via optimized tensor reshaping, plus compression, quantization, and pruning capabilities intended to reduce compute and data movement.
For data center engineers, the interesting part is NeuPro Ultra’s stated target: “AI infrastructure applications,” including “complex large language model inference in cloud and data centers,” and integration into “intelligent memory, connectivity and data-processing silicon.” That’s consistent with where inference bottlenecks often land at scale: memory bandwidth, KV cache growth, and the energy cost of shuttling data between compute, memory, and networks. But the practical question is how these blocks show up in real systems, since this is licensable IP and not a shipping server product.
Ceva says NeuPro Ultra supports multi-core processing, multi-user operation, high memory bandwidth, advanced memory management, and the software frameworks used in cloud and data center deployments. It also describes using the same neural processing technology inside memory and connectivity subsystems to reduce “unnecessary transfers” by processing data closer to where it’s stored.
“Working closely with the more than 15 customers who have licensed the first generation of NeuPro has shown us that successful AI products depend on the right balance of compute, memory, bandwidth, power and cost for each use case,” said Yaron Galitzky, Executive Vice President, AI Division at Ceva.
The next-generation NeuPro architecture is available now to select early-access customers, with general availability planned for the second quarter of 2027.
Source: Ceva
