Rockville – October 28, 2026 -- Ceva (NASDAQ: CEVA) has unveiled a next-generation NeuPro NPU portfolio that delivers up to 2X higher AI throughput per mm2 and 50% better energy efficiency than the first generation, and extends the company's neural processing IP into cloud and data center silicon for the first time.

Four NPUs share one architecture and one SDK

The portfolio comprises NeuPro Nano, Prime, Max and Ultra, all built on a common core architecture and a single unified software development kit. Chip designers can move between cores as requirements change, or combine several within one SoC and assign each workload to the best-suited engine. Ceva says this preserves software investment across product generations.

• NeuPro Nano: always-on sensing and entry-level perception in highly constrained devices.
• NeuPro Prime: mainstream CNN and transformer workloads.
• NeuPro Max: advanced edge AI, including high-resolution perception, multimodal models, on-device assistants, robotics and agentic AI.
• NeuPro Ultra: large language model inference in cloud and data centers, plus intelligent memory, connectivity and data-processing silicon.

FP4 and FP8 support targets memory and data-movement costs

The architecture adds FP4 and FP8 mixed-precision numerics to cut memory and data-movement requirements. It also adds higher memory bandwidth, dynamic shapes and paged KV cache for longer contexts and memory-intensive transformer workloads. Higher MAC utilization through optimized tensor reshaping is intended to raise inference performance from the same compute resources.

Supported network types include Mixture-of-Experts and linear attention. The cores also run pre-quantized models, so customers can deploy published checkpoints without a separate quantization and calibration step.

NeuPro Ultra moves Ceva into AI infrastructure

NeuPro Ultra offers multi-core processing, multi-user operation and advanced memory management for cloud and data center inference. It can also be embedded in memory and connectivity subsystems, where it applies neural processing to data movement rather than to the model itself, reducing unnecessary transfers. Ceva cites an ABI Research forecast of a 42% CAGR in cloud AI inference capacity through 2035.

CEO Amir Panush said a common NPU architecture lets customers "build once and deploy across their roadmap." He described NeuPro Ultra as an important step in Ceva's strategy to put neural processing into a broader range of semiconductor products.

More than 15 first-generation licensees informed the redesign

Yaron Galitzky, Executive Vice President of Ceva's AI Division, said work with more than 15 customers that licensed the first-generation NeuPro shaped the design. The focus was the balance of compute, memory, bandwidth, power and cost for each use case.

Early access opens now; general availability set for Q2 2027

The architecture is available now to select early-access customers. General availability is planned for the second quarter of 2027. Ceva says leading semiconductor and system companies are already evaluating and integrating the technology across advanced edge, Physical AI and AI infrastructure applications.

Ceva will present the portfolio and its Physical AI strategy at technology symposiums in Hsinchu, Taiwan, on November 2 and in Shanghai, China, on November 5.