GPU-driven exascale computing opens new and exciting opportunities for biophysical simulations, including the ability to address problems with greater length and time scales and to create more accurate microscopic models. Artificial intelligence (AI) methods have become indispensable tools that facilitate these large-scale simulations through model building, high-speed surrogate models that aid simulations, and analysis of the massive datasets produced.   

Both AI and simulations can be accelerated by exploiting mixed-precision calculations [1,2]. These accelerate tensor and matrix core operations, reducing memory and data transmission loads, as well as the energy consumption. In turn,  lower-precision computing has altered GPU hardware by allocating a greater portion of its resources to single- or lower-precision operations. Although there is extensive literature on mixed-precision calculations in applied mathematics—with topics ranging from linear algebra to computational fluid dynamics, FFTs, molecular dynamics, multigrid, and more—these topics are not always familiar to the MD community. To fully exploit increased speed and reduced memory requirements, extensive background work is needed to assess the impact of reduced-precision models on each portion of an algorithm. For example, what effect does reduced precision have on the dynamics and thermodynamics of a complex biological system, such as a protein? 

Another technology at the forefront of high-performance computing is quantum computing, a rapidly developing field [3.4]. Quantum computing has the potential to significantly speed up certain common computational tasks, such as sampling many-electron systems accurately, optimization, and machine learning. Although quantum computing requires significant infrastructure, including cryogenics and lasers, the computing itself is expected to require much less energy.