Mathieu Luisier

Afif Boudaoud, Lisa Gaedke-Merzhäuser, Alexandros Nikolaos Ziogas et al.

Spatio-temporal Bayesian inference drives environmental and health sciences using latent Gaussian models. Integrated Nested Laplace Approximations (INLA) enable inference for these models at HPC scale but rely on derivative-based optimization over $d$ hyperparameters. State-of-the-art INLA implementations approximate derivatives via central finite differences (FD), requiring $2d{+}1$ evaluations. These evaluations are embarrassingly parallel, but total work and energy grow with $d$, limiting time-to-solution under fixed budgets. Reverse-mode automatic differentiation (AD) computes exact gradients independently of $d$, but its efficient application to INLA's structured-sparse kernels is an open challenge. We present ADELIA, the first AD-enabled INLA implementation with a structure-exploiting multi-GPU backward pass leveraging model sparsity. We evaluate ADELIA on ten benchmark models, including real-world air-pollution monitoring. We achieve $4.2$--$7.9\times$ per-gradient speedups and reliable convergence on production-scale models with up to 1.9M latent variables, where FD struggles. Even when scaled to 16--32 GPUs to match ADELIA's wall-clock time, FD consumes $5$--$8\times$ more energy.

Denghui Lu, Alexander Maeder, Mathieu Luisier et al.

Modern GPUs devote an increasing silicon budget to low-precision matrix-multiplication units, widening the precision-throughput gap for scientific computing workloads. Ozaki Schemes I and II offer an alternative by reconstructing high-precision general matrix multiplication (GEMM) from low-precision operations, yet existing implementations leave substantial performance untapped. In particular, intermediate results are repeatedly materialized in global memory, making data movement the dominant bottleneck. We present EmuGEMM, fused integer Tensor Core kernels for NVIDIA Hopper and Blackwell GPUs that eliminate redundant memory round-trips in both Ozaki schemes. Using Scheme I, EmuGEMM sustains up to 1,639 Top/s on Hopper (83% of INT8 peak) and 3,654 Top/s on Blackwell (81%). For large matrices, EmuGEMM surpasses cuBLAS TF32 throughput by up to 1.4x on Hopper and 1.7x on Blackwell, at comparable accuracy. Using Scheme II, EmuGEMM extends to complex arithmetic and outperforms cuBLAS ZGEMM by up to 2.3x on Hopper and 5.5x on Blackwell.

Manasa Kaniselvan, Alexander Maeder, Chen Hao Xia et al.

Equivariant Graph Neural Networks (eGNNs) trained on density-functional theory (DFT) data can potentially perform electronic structure prediction at unprecedented scales, enabling investigation of the electronic properties of materials with extended defects, interfaces, or exhibiting disordered phases. However, as interactions between atomic orbitals typically extend over 10+ angstroms, the graph representations required for this task tend to be densely connected, and the memory requirements to perform training and inference on these large structures can exceed the limits of modern GPUs. Here we present a distributed eGNN implementation which leverages direct GPU communication and introduce a partitioning strategy of the input graph to reduce the number of embedding exchanges between GPUs. Our implementation shows strong scaling up to 128 GPUs, and weak scaling up to 512 GPUs with 87% parallel efficiency for structures with 3,000 to 190,000 atoms on the Alps supercomputer.