Custom deep learning processor
PFN is developing the MN-Core™ accelerator to speed up training of deep learning models. MN-Core is a dedicated accelerator optimized for matrix computations needed for deep learning, and is expected to achieve a world-class energy efficiency of 1 TFLOPS/W (half precision). By focusing on the functions required for deep learning, the dedicated chip can boost effective performance in deep learning as well as reduce costs.
We started operating MN-3, the first MN-Core-powered computer cluster with over 1,000 nodes, in May 2020 on a trial basis. Our goal is to increase MN-3’s calculation speed to 2 EFLOPS.
Optimized for the training phase in deep learning
Extremely densely integrated matrix arithmetic units
MN-Core-powered cluster MN-3 started operation in May 2020 on a trial basis
PFN, IIJ and JAIST to Launch Joint Research Project on Ultra-High-Efficiency AI Computing Infrastructure