Year: 2026
Type: dissertation
Abstract: Recent years have seen the widespread adoption of Machine Learning (ML) algorithms across a vast range of applications and scales. In order to execute these algorithms efficiently, hardware vendors have developed specialized hardware accelerators, called AI hardware. AI hardware is now present in many contemporary systems, including smartphones, laptops, servers, and supercomputers. Its widespread availability raises a natural question: can AI hardware also be used effectively for applications beyond ML?Scientific computing workloads are a promising target for this question. Many such workloads require high computational throughput, high memory bandwidth, and energy efficiency. For example, solving partial differential equations (PDEs) efficiently requires all of these properties. However, while these properties characterize modern AI hardware, such hardware is not designed for this purpose. AI hardware typically supports low-precision arithmetic, and the computational patterns it can execute efficiently are limited. This thesis explores whether AI hardware can be used for scientific computing, with a particular focus on stencil computations, linear system and PDE solvers. The central challenge is to express numerical methods in a form that can be compiled and executed efficiently by AI hardware toolchains, while still preserving the accuracy and performance requirements of scientific computing.AI hardware is usually targeted through frameworks such as PyTorch or TensorFlow, which allow computations to be expressed as computational graphs. These graphs are then compiled and executed using vendor-specific toolchains. In Paper I, we investigate this graph-based approach for describing PDE solvers, starting with linear PDEs with constant coefficients. Paper II expands on this result by introducing Thalassa, a transpiler that takes symbolic PDE systems together with a discretization scheme and generates PyTorch graphs implementing the corresponding solver. Papers I and II reveal two major limitations of AI hardware for PDE solving. The first is the lack of direct support for FP64 and higher precisions. The second is lack of support for certain patterns, which translates to increased memory movement during solving.The second part of this thesis addresses these limitations. For precision, we emulate high-precision arithmetic using low-precision operations. In particular, we adapt and extend the Ozaki scheme in Paper III to stencil computations. This scheme was originally developed for high-precision matrix multiplication. Our method can achieve arbitrarily high-precision results using INT8 primitives, which mitigates the precision problem. In addition, it can outperform the solvers on which it is based, mainly when the implemented stencil is FP64-compute bound on the target architecture. Paper IV generalizes this method further, and extends our algorithm to a practical application. It introduces OASIS, an INT8 Ozaki-based adaptive-precision solver that uses the representation’s flexibility to adjust the number of slices used in each computation. By doing so, OASIS can trade precision for performance in a controlled way. The results demonstrate that apart from being effective on isolated stencil kernels, our methods can also support realistic high-precision scientific workloads on low-precision AI hardware.
Author: Boulasikis, Michail
Institution: Lund University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225522824
