Modern GPUs feature specialized hardware for low-precision floating-point arithmetic to accelerate compute-intensive workloads that do not require high numerical accuracy, such as those from artificial intelligence. However, despite the significant gains in computational throughput, memory bandwidth utilization, and energy efficiency, integrating low-precision formats into scientific applications remains difficult. We introduce Kernel Float , a header-only C++ library that simplifies the development of portable mixed-precision GPU kernels. Kernel Float provides a generic vector type, a unified interface for common mathematical operations, and fast approximations for low-precision transcendental functions that lack native hardware support. To demonstrate the potential of mixed-precision computing unlocked by our library, we integrated Kernel Float into nine GPU kernels from various domains. Our evaluation on Nvidia A100 and AMD MI250X GPUs shows performance improvements of up to \(12\times\) over double precision, while reducing source code length by up to 50% compared to handwritten kernels and having negligible runtime overhead. Our results further show that mixed-precision performance depends not only on choosing appropriate data types but also on tuning traditional optimization parameters (e.g., block size and vector width) and, when relevant, even domain-specific parameters.

Reduced-precision floating-point arithmetic has become increasingly important in GPU applications for AI and HPC, as it can deliver substantial speedups while reducing energy consumption and memory footprint. However, choosing the appropriate data formats brings a challenging tuning problem: precision parameters must be chosen to maximize performance while preserving numerical accuracy. At the same time, GPU kernels typically expose additional tunable optimization parameters, such as block size, tiling strategy, and vector width. The combination of these two kinds of parameters results in a complex trade-off between accuracy and performance, making manual exploration of the resulting design space time-consuming. In this work, we present anaccuracy-awareextension to the open-sourceKernel Tunerframework, enabling automatic tuning of floating-point precision parameters alongside conventional code-optimization parameters. We evaluate our accuracy-aware tuning solution on both Nvidia and AMD GPUs using a variety of kernels. Our results show speedups of up to$12{\times }$over double precision, demonstrate how Kernel Tuner's built-in search strategies are effective for accuracy-aware tuning, and show that our approach can be extended to other optimization objectives, such as memory footprint or energy efficiency. Moreover, we highlight that jointly tuning accuracy- and performance-affecting parameters outperforms isolated approaches in finding the best-performing configurations, despite significantly expanding the optimization space. This unified approach enables developers to trade accuracy for throughput systematically, enabling broader adoption of mixed-precision computing in scientific and industrial applications.

The GPU programming model is primarily aimed at the development of applications that run one GPU. However, this limits the scalability of GPU code to the capabilities of a single GPU in terms of compute power and memory capacity. To scale GPU applications further, a great engineering effort is typically required: work and data must be divided over multiple GPUs by hand, possibly in multiple nodes, and data must be manually spilled from GPU memory to higher-level memories. We present Lightning: a framework that follows the common GPU programming paradigm but enables scaling to large problems with ease. Lightning supports multi-GPU execution of GPU kernels, even across multiple nodes, and seamlessly spills data to higher-level memories (main memory and disk). Existing CUDA kernels can easily be adapted for use in Lightning, with data access annotations on these kernels allowing Lightning to infer their data requirements and the dependencies between subsequent kernel launches. Lightning efficiently distributes the work/data across GPUs and maximizes efficiency by overlapping scheduling, data movement, and kernel execution when possible. We present the design and implementation of Lightning, as well as experimental results on up to 32 GPUs for eight benchmarks and one real-world application. Evaluation shows excellent performance and scalability, such as a speedup of 57.2 x over the CPU using Lighting with 16 GPUs over 4 nodes and 80 GB of data, far beyond the memory capacity of one GPU.

All-pairs compute problems apply a user-defined function to each combination of two items of a given data set. Although these problems present an abundance of parallelism, data reuse must be exploited to achieve good performance. Several researchers considered this problem, either resorting to partial replication with static work distribution or dynamic scheduling with full replication. In contrast, we present a solution that relies on hierarchical multi-level software-based caches to maximize data reuse at each level in the distributed memory hierarchy, combined with a divide-and-conquer approach to exploit data locality, hierarchical work-stealing to dynamically balance the workload, and asynchronous processing to maximize resource utilization. We evaluate our solution using three real-world applications, from digital forensics, localization microscopy, and bioinformatics, on different platforms, from desktop machine to a supercomputer. Results shows excellent efficiency and scalability when scaling to 96 GPUs, even obtaining super-linear speedups due to a distributed cache.

Studying the movements of crowds is important for understanding and predicting the behavior of large groups of people. When analyzing crowds, one is often interested in the long-term macro-level motions of the crowd as a whole, as opposed to the micro-level short-term movements of individuals. A high-level representation of these motions is thus desirable. In this work, we present a scalable method for detection of crowd motion patterns, i.e., spatial areas describing the dominant motions within crowds. For measuring crowd movements, we propose a fast, scalable, and low-cost method based on proximity graphs. For analyzing crowd movements, we utilize a three-stage pipeline: (1) represents the behavior of each person at each moment in time as a low-dimensional data point, (2) cluster these data points based on spatial relations, and (3) concatenate these clusters based on temporal relations. Experiments on synthetic datasets reveals our method can handle various scenarios including curved lanes and diverging flows. Evaluation on real-world datasets shows our method is able to extract useful motion patterns which could not be properly detected by existing methods. Overall, we see our work as an initial step towards rich pattern recognition.

In this paper we introduce LDBC Graphalytics, a new industrial-grade benchmark for graph analysis platforms. It consists of six deterministic algorithms, standard datasets, synthetic dataset generators, and reference output, that enable the objective comparison of graph analysis platforms. Its test harness produces deep metrics that quantify multiple kinds of system scalability, such as horizontal/vertical and weak/strong, and of robustness, such as failures and performance variability. The benchmark comes with open-source software for generating data and monitoring performance. We describe and analyze six implementations of the benchmark (three from the community, three from the industry), providing insights into the strengths and weaknesses of the platforms. Key to our contribution, vendors perform the tuning and benchmarking of their platforms.