Minisymposium

I will talk about IPUToolkit.jl, a package for running Julia code on the Intelligence Processing Unit (IPU), a massively parallel accelerator developed by Graphcore and powered by almost 1500 cores. I will show how Julia enables a high-degree of code reuse on this specialised hardware, using advanced packages such as DifferentialEquations.jl and Enzyme.jl.

PerfTest.jl is a Julia package conceived from the idea of bridging the gap between unit testing and architecture-agnostic performance testing. It brings a set of features that allow the user to set up performance regression unit tests from functional unit tests with minimal effort. An emphasis is made on letting the user create flexible test suites with classical performance models that can be applied across different machines.

We present a portable, GPU-accelerated implementation of a QR-based Singular Value algorithm in Julia, that allows code reproducibility across several different GPU vendors. Singular Value Decomposition (SVD) is a fundamental numerical tool in scientific computing and machine learning, providing optimal low-rank matrix approximations with applications ranging from dimensionality reduction to data compression and signal processing. Our implementation leverages Julia’s multiple dispatch and metaprogramming capabilities, integrating with the GPUArrays and KernelAbstractions frameworks to provide a unified type-, and hardware-agnostic API. It supports diverse GPU architectures and data types, including half precision and Apple Metal. We benchmark the algorithm against several state-of-the-art linear algebra libraries and confirm performance reproducibility through a unified API. We explore GPU kernel optimization through parameter tuning to enable efficient parallelism and improved memory locality. Performance results on multiple GPU backends and data types demonstrate scalability combined with reproducibility, highlighting Julia’s suitability for high-performance linear algebra in heterogeneous environments.

Julia is gaining traction in scientific computing, and at the Leibniz Supercomputing Centre (LRZ), we are exploring its potential on our high-performance computing (HPC) system, particularly on our Intel Ponte Vecchio GPUs of the SuperMUC-NG Phase 2 supercomputer. The Julia package KernelAbstraction.jl enables vendor-agnostic parallelization, allowing developers to write kernels that run efficiently on both multi-threaded CPUs and various GPU architectures with minimal modifications. This ability to write a single single-source, hardware-agnostic kernel bridges the gap between different hardware backends and enhances the reproducibility of both results and performance across diverse computing environments. To evaluate its real-world impact, we apply KernelAbstraction.jl to GaPSE.jl, a cosmology program that computes Two-Point Correlation Functions of galaxies including General Relativistic effects. GaPSE.jl needs to perform numerous nested integrals, which can be computationally expensive. By leveraging parallel execution on CPUs and GPUs, we aim to significantly accelerate these calculations, improving efficiency and scalability. In this talk, we share our experience developing and optimizing kernels with KernelAbstraction.jl, we benchmark Julia’s performance on HPC, and we show how reproducibility is ensured in a real-world application.