VSORA releases inference processor targeting efficient AI data center workloads

VSORA has announced a new AI inference processor targeting large-scale data center workloads and high-efficiency machine learning deployment. The chip, called Jotunn8, is designed for inference rather than training, addressing performance and energy challenges in running large AI models at scale.

The processor delivers a peak compute capacity of 3,200 teraflops, operating at more than 50% utilization while consuming about half the power of comparable devices used for similar workloads. This balance between throughput and power efficiency is intended to improve cost-effectiveness and sustainability for data centers expanding their AI capabilities.

Unlike GPUs optimized primarily for training, the Jotunn8 architecture focuses on real-time inference performance—executing pre-trained models for prediction, control, and decision tasks. Its design aims to minimize latency and energy loss while maintaining consistent utilization across varying workloads. The result is an inference engine that can be scaled across clusters for parallel deployment in high-demand computing environments.

Manufacturing of the chip is being carried out in partnership with global semiconductor fabrication companies to ensure quality, scalability, and access to advanced process technologies. These partnerships support both prototype and high-volume production, enabling integration into a range of data center and edge systems.