Ceva unveils NeuPro NPU family with 2x AI throughput per area
The next-generation NeuPro also claims 50% better energy efficiency than the first generation, with general availability planned for the second quarter of 2027.
Ceva announced its next-generation NeuPro NPU family on November 2, featuring an upgraded architecture and a unified software stack. The company claims the new design delivers up to 2x AI throughput per square millimeter and 50% better energy efficiency than the first generation, according to datanet.co.kr.
The architecture supports FP4 and FP8 numeric formats along with mixed-precision computation. Ceva says it is aimed at transformer workloads, with expanded memory bandwidth and features including dynamic tensor shapes and paged KV cache.
The family also supports Mixture-of-Experts and linear attention architectures, and can run pre-quantized model checkpoints without separate quantization or calibration.
Shared tooling across the family
A common SDK lets chip designers select engines across the family and reuse existing software assets when integrating multiple engines into a single SoC.
The next-generation NeuPro architecture is in early access with select customers and is scheduled for general availability in the second quarter of 2027.
Quick answers
When will the next-generation NeuPro be generally available?
Ceva says general availability is scheduled for the second quarter of 2027, with early access already under way for select customers.
What numeric formats does the new NeuPro support?
The architecture supports FP4 and FP8 formats as well as mixed-precision computation.
How much faster is the next-generation NeuPro?
Ceva claims up to 2x AI throughput per square millimeter and 50% better energy efficiency than the first generation.