Gerhard Wellein

Moritz Kreutzer, Georg Hager, Gerhard Wellein et al.

Sparse matrix-vector multiplication (spMVM) is the dominant operation in many sparse solvers. We investigate performance properties of spMVM with matrices of various sparsity patterns on the nVidia "Fermi" class of GPGPUs. A new "padded jagged diagonals storage" (pJDS) format is proposed which may substantially reduce the memory overhead intrinsic to the widespread ELLPACK-R scheme. In our test scenarios the pJDS format cuts the overall spMVM memory footprint on the GPGPU by up to 70%, and achieves 95% to 130% of the ELLPACK-R performance. Using a suitable performance model we identify performance bottlenecks on the node level that invalidate some types of matrix structures for efficient multi-GPGPU parallelization. For appropriate sparsity patterns we extend previous work on distributed-memory parallel spMVM to demonstrate a scalable hybrid MPI-GPGPU code, achieving efficient overlap of communication and computation.

Bole Ma, Jan Eitzinger, Harald Köstler et al.

Frontier LLMs increasingly decide what a query attends to with a sparse-attention indexer that picks a few KV-cache blocks per query: attention's unit is now a small, reusable chunk. Agentic workloads hammer it: many sub-agents query one large codebase, reusing the same blocks. When that corpus outgrows one GPU it is partitioned across instances, so a query and the blocks it selects often sit on different GPUs: answering it means attention across instances. The reflex of prior cross-instance KV systems is to move the cache: pull the selected blocks to the requester. Multi-head Latent Attention inverts the arithmetic, compressing each token's key and value into one narrow vector, so a routed query row is only ~1 KB, smaller than the chunk it attends; routing the query is then often cheaper than moving the cache. Which primitive wins, over which fabric and request shape, is uncharted, least of all on device-initiated RDMA that makes per-request cross-node transfers cheap. We characterize cross-instance MLA attention on a real multi-node H100 cluster, distilling two reusable artifacts: a topology-aware cost model (probe / transfer / compute / return / merge) and a closed-form route/fetch/local predicate, whose constants we measure on real IBGDA, where the model tracks batched round-trips to within ~7%. At decode it routes the query, trading the cost of moving the cache (a ~3 ms re-adaptation splice for a contiguous chunk, or a scattered gather under selection) for a tens-of-microsecond round trip, and picks the fabric by probe latency, not peak bandwidth. We instantiate the cost model and predicate for MLA, but neither is MLA-specific: they apply wherever compression or sparse selection shrinks attention to small chunks (DeepSeek-V3.2, V4, and GLM-5.1 today). Extending them to a new architecture requires measuring just two coefficients: the routed payload and fetch's move-the-cache cost.

Ayesha Afzal, Georg Hager, Gerhard Wellein et al.

This paper studies the utility of using data analytics and machine learning techniques for identifying, classifying, and characterizing the dynamics of large-scale parallel (MPI) programs. To this end, we run microbenchmarks and realistic proxy applications with the regular compute-communicate structure on two different supercomputing platforms and choose the per-process performance and MPI time per time step as relevant observables. Using principal component analysis, clustering techniques, correlation functions, and a new "phase space plot," we show how desynchronization patterns (or lack thereof) can be readily identified from a data set that is much smaller than a full MPI trace. Our methods also lead the way towards a more general classification of parallel program dynamics.

Christie Alappat, Jonas Thies, Georg Hager et al.

Sparse linear iterative solvers are essential for many large-scale simulations. Much of the runtime of these solvers is often spent in the implicit evaluation of matrix polynomials via a sequence of sparse matrix-vector products. A variety of approaches has been proposed to make these polynomial evaluations explicit (i.e., fix the coefficients), e.g., polynomial preconditioners or s-step Krylov methods. Furthermore, it is nowadays a popular practice to approximate triangular solves by a matrix polynomial to increase parallelism. Such algorithms allow to evaluate the polynomial using a so-called matrix power kernel (MPK), which computes the product between a power of a sparse matrix A and a dense vector x, or a related operation. Recently we have shown that using the level-based formulation of sparse matrix-vector multiplications in the Recursive Algebraic Coloring Engine (RACE) framework we can perform temporal cache blocking of MPK to increase its performance. In this work, we demonstrate the application of this cache-blocking optimization in sparse iterative solvers. By integrating the RACE library into the Trilinos framework, we demonstrate the speedups achieved in preconditioned) s-step GMRES, polynomial preconditioners, and algebraic multigrid (AMG). For MPK-dominated algorithms we achieve speedups of up to 3x on modern multi-core compute nodes. For algorithms with moderate contributions from subspace orthogonalization, the gain reduces significantly, which is often caused by the insufficient quality of the orthogonalization routines. Finally, we showcase the application of RACE-accelerated solvers in a real-world wind turbine simulation (Nalu-Wind) and highlight the new opportunities and perspectives opened up by RACE as a cache-blocking technique for MPK-enabled sparse solvers.

Aditya Ujeniya, Jan Eitzinger, Georg Hager et al.

Modern NVIDIA GPUs like the H100 (HBM2e) and H200 (HBM3e) share similar compute characteristics but differ significantly in memory interface technology and bandwidth. By isolating memory bandwidth as a key variable, the power distribution between the memory and Streaming Multiprocessors (SM) changes notably between the two architectures. In the era of energy-efficient computing, analyzing how these hardware characteristics impact performance per watt is critical. This study investigates how the H100 and H200 manage memory power consumption at various power-cap levels. By a regression analysis, we study the memory power limit and uncover outliers consuming more memory power. To evaluate efficiency, we employ compute-bound (DGEMM) and memory-bound (TheBandwidthBenchmark) workloads, representing the two extremes of the Roof\-line model. Our observations indicate that across varying power caps, the H100 remains the slightly better choice for strictly compute-bound workloads, whereas the H200 demonstrates superior efficiency for memory-bound applications.

Dane C. Lacey, Christie L. Alappat, Florian Lange et al.

Sparse matrix-vector products (SpMVs) are a bottleneck in many scientific codes. Due to the heavy strain on the main memory interface from loading the sparse matrix and the possibly irregular memory access pattern, SpMV typically exhibits low arithmetic intensity. Repeating these products multiple times with the same matrix is required in many algorithms. This so-called matrix power kernel (MPK) provides an opportunity for data reuse since the same matrix data is loaded from main memory multiple times, an opportunity that has only recently been exploited successfully with the Recursive Algebraic Coloring Engine (RACE). Using RACE, one considers a graph based formulation of the SpMV and employs s level-based implementation of SpMV for reuse of relevant matrix data. However, the underlying data dependencies have restricted the use of this concept to shared memory parallelization and thus to single compute nodes. Enabling cache blocking for distributed-memory parallelization of MPK is challenging due to the need for explicit communication and synchronization of data in neighboring levels. In this work, we propose and implement a flexible method that interleaves the cache-blocking capabilities of RACE with an MPI communication scheme that fulfills all data dependencies among processes. Compared to a "traditional" distributed memory parallel MPK, our new Distributed Level-Blocked MPK yields substantial speed-ups on modern Intel and AMD architectures across a wide range of sparse matrices from various scientific applications. Finally, we address a modern quantum physics problem to demonstrate the applicability of our method, achieving a speed-up of up to 4x on 832 cores of an Intel Sapphire Rapids cluster.

Bole Ma, Jan Eitzinger, Harald Koestler et al.

AlltoAll dispatch is the dominant bottleneck of MoE expert parallelism, and the interconnect community has responded with four families of mitigations: predictive sample placement, adaptive expert relayout, hierarchical collectives, and EP-aware topology. All four rest on two assumptions about the workload. The first is that routing imbalance is correctable by the system layer. The second is that the mock-token benchmarks evaluating them faithfully represent production routing. We introduce DODOCO to test both assumptions. We instrument five MoE checkpoints spanning five sequence-mixer designs (DeepSeek-V2-Lite MLA, DeepSeek-MoE-16B MHA, Qwen3-30B GQA, Nemotron-30B Mamba-2, Qwen3.5-35B GDN) under a 5 by 6 grid of data conditions plus a matched EP scan from 4 to 32 ranks on H100s; both assumptions fail. Scaling EP changes the per-expert max/mean token ratio by at most 5% within every architecture's measurable range: the straggler is intrinsic to the routing decision the model makes, not to how its experts land on ranks. Mock tokens overestimate routing Gini by up to a factor of 2.35 and fabricate a batch-size scaling trend that vanishes the moment real text replaces random IDs. A third pattern, unexpected, emerges from the same matrix: the five architectures cleave into two stable bands. MHA and Mamba-2 (data-resilient) drop to Gini 0.105 and 0.150 on wikitext. MLA and GDN (persistently concentrated) stay above 0.24 on every real-text condition and reach 0.29 to 0.38 on mock. GQA is the intermediate case. These bands, not the EP degree or the mock-data profile, are the right workload input to AlltoAll-aware interconnect and dispatch design.

Sebastian Wind, Jeta Sopa, Laurin Schmid et al.

Large language models (LLMs) demonstrate strong general reasoning and language understanding, yet their performance degrades in domains governed by strict formal rules, precise terminology, and legally binding structure. Tax law exemplifies these challenges, as correct answers require exact statutory citation, structured legal argumentation, and numerical accuracy under rigid grading schemes. We algorithmically generate SteuerEx, the first open benchmark derived from authentic German university tax law examinations. SteuerEx comprises 115 expert-validated examination questions spanning six core tax law domains and multiple academic levels, and employs a statement-level, partial-credit evaluation framework that closely mirrors real examination practice. We further present SteuerLLM, a domain-adapted LLM for German tax law trained on a large-scale synthetic dataset generated from authentic examination material using a controlled retrieval-augmented pipeline. SteuerLLM (28B parameters) consistently outperforms general-purpose instruction-tuned models of comparable size and, in several cases, substantially larger systems, demonstrating that domain-specific data and architectural adaptation are more decisive than parameter scale for performance on realistic legal reasoning tasks. All benchmark data, training datasets, model weights, and evaluation code are released openly to support reproducible research in domain-specific legal artificial intelligence. A web-based demo of SteuerLLM is available at https://steuerllm.i5.ai.fau.de.

Martin Mayr, Sebastian Wind, Lukas Schröder et al.

Artificial Intelligence (AI) workloads drive a rapid expansion of high-performance computing (HPC) infrastructures and increase their power and energy demands towards a critical level. AI benchmarks representing state-of-the art workloads and their understanding in the context of performance-energy trade-offs are critical to deploy efficient infrastructures and can guide energy efficiency measures, such as power capping. We introduce a benchmarking framework with popular deep learning applications from computer vision (image classification and generation) and large language models (continued pre-training and inference) implementing modern methods. Our performance analysis focuses on throughput rather than time to "completion", which is the standard metric in HPC. We analyse performance and energy efficiency under various power capping scenarios on NVIDIA H100, NVIDIA H200, and AMD MI300X GPUs. Our results reveal that no universal optimal power cap exists, as the efficiency peak varies across application types and GPU architectures. Interestingly, the two NVIDIA GPUs which mainly differ in their HBM configuration show qualitatively different performance-energy trade-offs. The developed benchmarking framework will be released as a public tool.

Bole Ma, Ayesha Afzal, Jan Eitzinger et al.

Power capping is the standard GPU energy lever in LLM serving, and it appears to work: throughput drops, power readings fall, and energy budgets are met. We show the appearance is illusory for the phase that dominates production serving: autoregressive decode. Across four attention paradigms -- GQA, MLA, Gated DeltaNet, and Mamba2 -- on NVIDIA H200, decode draws only 137--300\,W on a 700\,W GPU; no cap ever triggers, because memory-bound decode saturates HBM bandwidth rather than compute and leaves power headroom untouched. Firmware-initiated clock throttling compounds the illusion: these deviations can corrupt any throughput measurement that attributes them to the cap. SM clock locking dissolves both confounds. By targeting the lever that is actually on the critical path, clock locking Pareto-dominates power capping universally, recovering up to 32\% of decode energy at minimal throughput loss. We identify three architecture-dependent DVFS behavioural classes and characterise a common energy pattern across novel attention replacements: a heavy prefill cost recouped by efficient decode, eventually halving total request energy relative to GQA at production batch sizes.

Jan Laukemann, Julian Hammer, Johannes Hofmann et al.

An accurate prediction of scheduling and execution of instruction streams is a necessary prerequisite for predicting the in-core performance behavior of throughput-bound loop kernels on out-of-order processor architectures. Such predictions are an indispensable component of analytical performance models, such as the Roofline and the Execution-Cache-Memory (ECM) model, and allow a deep understanding of the performance-relevant interactions between hardware architecture and loop code. We present the Open Source Architecture Code Analyzer (OSACA), a static analysis tool for predicting the execution time of sequential loops comprising x86 instructions under the assumption of an infinite first-level cache and perfect out-of-order scheduling. We show the process of building a machine model from available documentation and semi-automatic benchmarking, and carry it out for the latest Intel Skylake and AMD Zen micro-architectures. To validate the constructed models, we apply them to several assembly kernels and compare runtime predictions with actual measurements. Finally we give an outlook on how the method may be generalized to new architectures.

Sebastian Wind, Tri-Thien Nguyen, Jeta Sopa et al.

Clinical LLMs are often scaled by increasing model size, context length, retrieval complexity, or inference-time compute, with the implicit expectation that higher accuracy implies safer behavior. This assumption is incomplete in medicine, where a few confident, high-risk, or evidence-contradicting errors can matter more than average benchmark performance. We introduce SaFE-Scale, a framework for measuring how clinical LLM safety changes across model scale, evidence quality, retrieval strategy, context exposure, and inference-time compute. To instantiate this framework, we introduce RadSaFE-200, a Radiology Safety-Focused Evaluation benchmark of 200 multiple-choice questions with clinician-defined clean evidence, conflict evidence, and option-level labels for high-risk error, unsafe answer, and evidence contradiction. We evaluated 34 locally deployed LLMs across six deployment conditions: closed-book prompting (zero-shot), clean evidence, conflict evidence, standard RAG, agentic RAG, and max-context prompting. Clean evidence produced the strongest improvement, increasing mean accuracy from 73.5% to 94.1%, while reducing high-risk error from 12.0% to 2.6%, contradiction from 12.7% to 2.3%, and dangerous overconfidence from 8.0% to 1.6%. Standard RAG and agentic RAG did not reproduce this safety profile: agentic RAG improved accuracy over standard RAG and reduced contradiction, but high-risk error and dangerous overconfidence remained elevated. Max-context prompting increased latency without closing the safety gap, and additional inference-time compute produced only limited gains. Worst-case analysis showed that clinically consequential errors concentrated in a small subset of questions. Clinical LLM safety is therefore not a passive consequence of scaling, but a deployment property shaped by evidence quality, retrieval design, context construction, and collective failure behavior.

Ayesha Afzal, Georg Hager, Gerhard Wellein

The escalating computational demands and energy footprint of GPU-accelerated computing systems complicate informed design and operational decisions. We present the first release of Wattlytics (https://wattlytics.netlify.app), an interactive, browser-based decision-support system. Unlike existing procurement-oriented calculators, Wattlytics uniquely integrates benchmark-driven GPU performance scaling, dynamic voltage and frequency scaling (DVFS)-aware piecewise power modeling, and multi-year total cost of ownership (TCO) analysis within a single interactive environment. Users can configure heterogeneous systems across contemporary GPU architectures (GH200, H100, L40S, L40, A40, A100, and L4), select representative scientific workloads (e.g., GROMACS, AMBER), and explore deployment scenarios under constraints such as energy prices, system lifetime, and frequency scaling. Wattlytics computes multidimensional decision metrics (TCO breakdown, work-per-TCO, power-per-TCO, and work-per-watt-per-TCO) and supports design-space exploration, what-if scenarios, sensitivity metrics (elasticity, Sobol indices, Monte Carlo) and collaborative features to guide realistic cluster design and procurement under uncertainty. We demonstrate selected scenarios comparing deployment strategies under different operational modes: ixed budget, fixed GPU count, fixed performance, and fixed power. Our case studies show that, under budget or energy constraints, optimally deployed energy-efficient GPUs can outperform higher-performance alternatives in overall cost-effectiveness. Wattlytics helps users explore the design parameter space and distinguish between cost- and risk-driving factors, turning HPC design into a well-informed and explainable decision-making process.

Sebastian Wind, Jeta Sopa, Daniel Truhn et al.

Clinical decision-making in radiology increasingly benefits from artificial intelligence (AI), particularly through large language models (LLMs). However, traditional retrieval-augmented generation (RAG) systems for radiology question answering (QA) typically rely on single-step retrieval, limiting their ability to handle complex clinical reasoning tasks. Here we propose radiology Retrieval and Reasoning (RaR), a multi-step retrieval and reasoning framework designed to improve diagnostic accuracy, factual consistency, and clinical reliability of LLMs in radiology question answering. We evaluated 25 LLMs spanning diverse architectures, parameter scales (0.5B to >670B), and training paradigms (general-purpose, reasoning-optimized, clinically fine-tuned), using 104 expert-curated radiology questions from previously established RSNA-RadioQA and ExtendedQA datasets. To assess generalizability, we additionally tested on an unseen internal dataset of 65 real-world radiology board examination questions. RaR significantly improved mean diagnostic accuracy over zero-shot prompting and conventional online RAG. The greatest gains occurred in small-scale models, while very large models (>200B parameters) demonstrated minimal changes (<2% improvement). Additionally, RaR retrieval reduced hallucinations (mean 9.4%) and retrieved clinically relevant context in 46% of cases, substantially aiding factual grounding. Even clinically fine-tuned models showed gains from RaR (e.g., MedGemma-27B), indicating that retrieval remains beneficial despite embedded domain knowledge. These results highlight the potential of RaR to enhance factuality and diagnostic accuracy in radiology QA, warranting future studies to validate their clinical utility. All datasets, code, and the full RaR framework are publicly available to support open research and clinical translation.

Dimosthenis Pasadakis, Christie Louis Alappat, Olaf Schenk et al.

Nonlinear reformulations of the spectral clustering method have gained a lot of recent attention due to their increased numerical benefits and their solid mathematical background. We present a novel direct multiway spectral clustering algorithm in the $p$-norm, for $p \in (1, 2]$. The problem of computing multiple eigenvectors of the graph $p$-Laplacian, a nonlinear generalization of the standard graph Laplacian, is recasted as an unconstrained minimization problem on a Grassmann manifold. The value of $p$ is reduced in a pseudocontinuous manner, promoting sparser solution vectors that correspond to optimal graph cuts as $p$ approaches one. Monitoring the monotonic decrease of the balanced graph cuts guarantees that we obtain the best available solution from the $p$-levels considered. We demonstrate the effectiveness and accuracy of our algorithm in various artificial test-cases. Our numerical examples and comparative results with various state-of-the-art clustering methods indicate that the proposed method obtains high quality clusters both in terms of balanced graph cut metrics and in terms of the accuracy of the labelling assignment. Furthermore, we conduct studies for the classification of facial images and handwritten characters to demonstrate the applicability in real-world datasets.

Johannes Hofmann, Jan Treibig, Georg Hager et al.

We examine the Xeon Phi, which is based on Intel's Many Integrated Cores architecture, for its suitability to run the FDK algorithm--the most commonly used algorithm to perform the 3D image reconstruction in cone-beam computed tomography. We study the challenges of efficiently parallelizing the application and means to enable sensible data sharing between threads despite the lack of a shared last level cache. Apart from parallelization, SIMD vectorization is critical for good performance on the Xeon Phi; we perform various micro-benchmarks to investigate the platform's new set of vector instructions and put a special emphasis on the newly introduced vector gather capability. We refine a previous performance model for the application and adapt it for the Xeon Phi to validate the performance of our optimized hand-written assembly implementation, as well as the performance of several different auto-vectorization approaches.