Spotlight: University of Tokyo Uses NVIDIA Grace Hopper for Groundbreaking Energy-Efficient Seismic Research

Supercomputers are the engines of groundbreaking discoveries. From predicting extreme weather to advancing disease research and designing safer, more efficient infrastructures, these machines simulate complex systems that are impractical to test in the real world due to their size, cost, and material requirements.

Since the introduction of the GPU in 1999, NVIDIA has continually pushed the boundaries of accelerated computing, an approach that uses specialized hardware to dramatically speed up work by bundling frequently occurring tasks in parallel processing. This efficiency enables supercomputers to handle unprecedented computational challenges while consuming less energy per task.

Today, NVIDIA-powered systems lead the charge in energy-efficient supercomputing, with eight of the top 10 placements on the Green500—an industry benchmark for energy-efficient supercomputers. The JEDI system at Germany’s Jülich Supercomputing Center, powered by NVIDIA Grace Hopper, exemplifies this progress, achieving a staggering efficiency of 72.7 gigaflops per watt.

Advancing seismic safety with the University of Tokyo

This drive for energy efficiency is exemplified in the recent research effort between NVIDIA and the University of Tokyo. Situated in one of the most seismically active regions in the world, Japan requires cutting-edge research to mitigate the impact of earthquakes.

The Earthquake Research Institute at the University of Tokyo has used the NVIDIA Grace Hopper Superchip to accelerate its seismic simulations, achieving groundbreaking results.

The tightly coupled CPU-GPU architecture of the NVIDIA GH200 Grace Hopper Superchip enables an improvement in simulation performance of 86x, with 32x greater energy efficiency compared to traditional methods. Advanced memory optimization ensures faster and more complex computations, empowering researchers to model earthquake-resistant infrastructure with unmatched precision. By leveraging data-driven predictions and heterogeneous computing systems, the university is helping transform seismic research into a more efficient endeavor.

Award-winning research: Heterogeneous computing for seismic modeling

These breakthroughs were presented at WACCPD 2024, where the University of Tokyo showcased how NVIDIA hardware-software synergy enables solutions to dynamic, large-scale problems. In seismic modeling, the innovative architecture of Grace Hopper accelerates time-evolution equation-based simulations by combining high-performance GPUs for intensive calculations with the memory capacity of CPUs for predictive algorithms.

Beyond seismic research, this approach has the potential to reshape energy efficiency in high-performance computing (HPC).

This groundbreaking work developed a new CPU-GPU heterogeneous computing method for repeatedly solving time-evolution partial differential equation (PDE) problems with guaranteed accuracy. This new method achieves a very short time-to-solution (TTS) with low energy-to-solution by leveraging both the large memory capacity of the CPU and the high computing performance of the GPU. When scaled up to the Alps supercomputer, the method was 51.6x faster than using only the CPU and 6.98x faster than the GPU, while achieving an impressive 94.3% efficiency across 1,920 compute nodes.

Bar chart comparison of NVIDIA GH200 system performance using both CPU and GPU with baseline method using only CPU or GPU. — *Figure 1. Performance comparison of the newly-developed PDE solver. Using both CPU and GPU in the NVIDIA GH200 Grace Hopper Superchip realizes maximum performance*

The simulations involve stepping forward in time, where the solution (the seismic activity, for example) at each point in time is calculated by solving a system of equations. In turn, at each timestep multiple iterations are required, where the solution converges to a suitably accurate result. The calculations at each iteration are very intense and require the computational power of GPUs.

Therefore, there are two aspects that determine the overall TTS (and corresponding energy to solution): how fast each iteration can be calculated at each timestep, and how many iterations are required at each timestep. It is this combination that enables Grace Hopper architecture to really excel: the Hopper GPU performs each calculation exceedingly quickly, while the large memory capacity of the Grace CPU stores a history of prior results to inform the predictor, allowing the number of iterations to be vastly reduced.

This data-driven method is well suited to the NVIDIA GH200 Grace Hopper Superchip. The number of iterations required to converge depends on how good the initial “guess” of the solution is. The novelty of this work involves data from previous timesteps being used to maximize the accuracy of each initial solution, thus reducing the number of iterations required. This method requires a combination of capabilities:

The large memory capacity of the NVIDIA Grace CPU to hold the required previous-timestep data,
The computational ability of Hopper to solve the equations at each iteration,
The fast interconnect to allow the data-driven results from Grace CPU to optimize the number of solver iterations on the Hopper GPU.

Diagram with Grace CPU on the left and Hopper GPU on the right, with arrows between predictors (left) and solvers (right) showing the ping-pong approach where the Grace CPU uses previous timestep results to predict the solution of the following step. Once completed the prediction is transferred via C2C to the GPU for the computational heavy lifting. This approach eliminates about ⅔ of the iteration time. — *Figure 2. The data-driven method solves two problem sets simultaneously by using the high bandwidth CPU-GPU connection on GH200 to communicate predictions from the CPU to the solver on the GPU*

A comparison of the iteration time history of relative error of the numerical method. Two sets of lines are shown, one for the baseline method which requires about 150 iterations to get to machine zero where the data-driven predictor method requires about 50 iterations because it starts from a much lower solver error. — *Figure 3. Data-driven predictor running on the Grace CPU reduces iterations from about 150 to 50*

The benefits are shown in Figure 1, where it can be seen that the time-to-solution is reduced by 86x, compared to only using the CPU, or 9x compared to only using the GPU. The corresponding energy reductions are 32x (compared to CPU) and 7x (compared to GPU alone).

Enabling energy efficient supercomputing

The University of Tokyo’s seismic research breakthroughs highlight the transformative power of energy-efficient supercomputing. By harnessing the NVIDIA Grace Hopper Superchip, researchers are not only accelerating seismic simulations by unprecedented margins but also drastically reducing energy consumption—a crucial step in sustainable computing.

This collaboration demonstrates how cutting-edge technology can address urgent global challenges, like earthquake safety, while setting new benchmarks for performance and energy efficiency. As these innovations continue to scale across industries, they pave the way for a future where high-performance computing drives both scientific progress and environmental responsibility.

At NVIDIA GTC 2025, Kohei Fujita, associate professor at the University of Tokyo Earthquake Research Institute, will discuss this breakthrough research in the session, Implement Accelerated PDE-Based Time-History Simulation by Data-Driven Methods on Strongly Coupled CPU-GPU systems [S72925].