NVIDIA adopts Shibaura Institute of Technology computing method

NVIDIA has integrated the ‘Ozaki Scheme II’ computing method into the latest version of its GPU development platform, CUDA Toolkit 13.4. Developed by Professor Katsuhisa Ozaki and his team at the Shibaura Institute of Technology, this technique significantly accelerates double-precision calculations by mathematically combining multiple low-precision calculations typically used for AI.

Unlike previous methods, Ozaki Scheme II operates as software on existing NVIDIA Ampere-generation GPUs and later, requiring no additional hardware modifications. This allows scientific simulations—such as weather forecasting and earthquake resistance design—to utilize the high-performance low-precision arithmetic units optimized for AI, effectively reclaiming high-precision performance without sacrificing accuracy.

On NVIDIA B200 GPUs, the method has achieved up to 175 TFLOPS for real double-precision matrix multiplication (DGEMM) and 295 TFLOPS for complex numbers (ZGEMM). The technology is expected to be utilized in major computational infrastructures, including the ‘Fugaku NEXT’ supercomputer being developed by RIKEN and various scientific research programs under the U.S. Department of Energy.