Younggeun Choi, Junyoung Park, Sang Min Lee, Jeseung Yeon, Minho Kim, Changjae Park, Byeongwook Bae, Hyunmin Jeong, Hanjoon Kim, June Paik, Nuno P. Lopes, Sungjoo Yoo
Modern AI workloads require architectures capable of efficiently managing diverse tensor contraction patterns. Traditional approaches based on fixed-size matrix multiplications often fall short in scalability and flexibility.
RNGD (pronounced "Renegade"), a second-generation tensor contraction processor (TCP), introduces an innovative architecture designed to exploit the parallelism and data locality inherent in tensor computations. Its coarse-grained processing elements (PEs) can operate as a unified large-scale unit or as multiple independent units, providing flexibility for various tensor shapes. Key innovations, such as a circuit switch-based fetch network, input broadcasting, and buffer-based reuse mechanisms, further enhance computational efficiency.
RNGD represents a significant advancement in processor architecture, delivering optimized performance and energy efficiency for sustainable computation of next-generation AI workloads.
Y. Choi, J. Park, S. M. Lee, J. Yeon, M. Kim, C. Park, B. Bae, H. Jeong, H. Kim, J. Paik, N. P. Lopes, S. Yoo. FuriosaAI RNGD: A Tensor Contraction Processor for Sustainable AI Computing. IEEE Micro, 2025.
@article{tcp-micro25,
title = {{FuriosaAI} {RNGD}: A Tensor Contraction Processor for Sustainable {AI} Computing},
author = {Younggeun Choi and Junyoung Park and Sang Min Lee and Jeseung Yeon and Minho Kim and Changjae Park and Byeongwook Bae and Hyunmin Jeong and Hanjoon Kim and June Paik and Nuno P. Lopes and Sungjoo Yoo},
journal = {IEEE Micro},
publisher = {IEEE},
doi = {10.1109/MM.2025.3551880},
year = 2025
}
© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.