Preferred Networks and Tenstorrent partner to develop generative AI processors

Preferred Networks (PFN) and US-based Tenstorrent have announced a joint development project to create high-performance, high-efficiency systems for generative AI training and inference. Under the agreement, Tenstorrent will lead CPU development, while PFN will focus on the AI processor, aiming to compete with the current industry standard held by NVIDIA GPUs.

The collaboration will utilize advanced packaging and optical-electrical convergence technology to enhance data transfer efficiency and reduce power consumption. The project has been selected for research and development by the New Energy and Industrial Technology Development Organization (NEDO) in Japan, with a goal to commercialize the results following research through fiscal year 2031.

Tenstorrent, led by chip designer Jim Keller, specializes in RISC-V based processors. PFN is known for its MN-Core series, which has previously achieved top rankings in power efficiency for supercomputing. The partnership aims to leverage these strengths to provide high-speed inference capabilities for large language models (LLMs) within server and desktop-sized form factors.