Cristina Silvano

Cristina Silvano, Daniele Ielmini, Fabrizio Ferrandi et al.

Recent trends in deep learning (DL) have made hardware accelerators essential for various high-performance computing (HPC) applications, including image classification, computer vision, and speech recognition. This survey summarizes and classifies the most recent developments in DL accelerators, focusing on their role in meeting the performance demands of HPC applications. We explore cutting-edge approaches to DL acceleration, covering not only GPU- and TPU-based platforms but also specialized hardware such as FPGA- and ASIC-based accelerators, Neural Processing Units, open hardware RISC-V-based accelerators, and co-processors. This survey also describes accelerators leveraging emerging memory technologies and computing paradigms, including 3D-stacked Processor-In-Memory, non-volatile memories like Resistive RAM and Phase Change Memories used for in-memory computing, as well as Neuromorphic Processing Units, and Multi-Chip Module-based accelerators. Furthermore, we provide insights into emerging quantum-based accelerators and photonics. Finally, this survey categorizes the most influential architectures and technologies from recent years, offering readers a comprehensive perspective on the rapidly evolving field of deep learning acceleration.

Serena Curzel, Fabrizio Ferrandi, Leandro Fiorin et al.

Given their increasing size and complexity, the need for efficient execution of deep neural networks has become increasingly pressing in the design of heterogeneous High-Performance Computing (HPC) and edge platforms, leading to a wide variety of proposals for specialized deep learning architectures and hardware accelerators. The design of such architectures and accelerators requires a multidisciplinary approach combining expertise from several areas, from machine learning to computer architecture, low-level hardware design, and approximate computing. Several methodologies and tools have been proposed to improve the process of designing accelerators for deep learning, aimed at maximizing parallelism and minimizing data movement to achieve high performance and energy efficiency. This paper critically reviews influential tools and design methodologies for Deep Learning accelerators, offering a wide perspective in this rapidly evolving field. This work complements surveys on architectures and accelerators by covering hardware-software co-design, automated synthesis, domain-specific compilers, design space exploration, modeling, and simulation, providing insights into technical challenges and open research directions.

Marco Ronzani, Cristina Silvano

Hypergraph partitioning is a recurring NP-hard problem in engineering; its efficient solution at scale hinges on parallelism. This work proposes a GPU-centric algorithm for multi-level hypergraph partitioning aimed at a specific set of problem constraints: limited size and distinct inbound hyperedges per partition. Manipulating hypergraphs requires deeply nested traversals and concurrent decision-making; our constraints impose further set operations amidst that. In turn, we design algorithms around the GPU's hierarchical parallelism and our problem's specifics. When forming partitions, we materialize the hypergraph's incidence structure and unique neighborhoods in memory to exploit set sparsity and batch node-pairing scores in shared memory. Upon refining partitions, we chain node moves into improving paths and cycles, checking their validity via cumulative set size variations reduced in parallel over moves. Thus, our dominant kernels exhibit a span linear in local hypergraph parameters. Results show an average 380x speedup and a 1.2-2.0x reduction in connectivity compared to a sequential multi-level partitioner. With minor changes, we also support k-way balanced partitioning, running 5x faster than CPU methods with a ~5% quality loss for k=2, outperforming an existing GPU partitioner at comparable runtime, with no measurable overhead from the added constraints handling logic.

Marco Ronzani, Cristina Silvano

Executing Spiking Neural Networks (SNNs) on neuromorphic hardware poses the problem of mapping neurons to cores. SNNs operate by propagating spikes between neurons that form a graph through synapses. Neuromorphic hardware mimics them through a network-on-chip, transmitting spikes, and a mesh of cores, each managing several neurons. Its operational cost is tied to spike movement and active cores. A mapping comprises two tasks: partitioning the SNN's graph to fit inside cores and placement of each partition on the hardware mesh. Both are NP-hard problems, and as SNNs and hardware scale towards billions of neurons, they become increasingly difficult to tackle effectively. In this work, we propose to raise the abstraction of SNNs from graphs to hypergraphs, redesigning mapping techniques accordingly. The resulting model faithfully captures the replication of spikes inside cores by exposing the notion of hyperedge co-membership between neurons. We further show that the overlap and locality of hyperedges strongly correlate with high-quality mappings, making these properties instrumental in devising mapping algorithms. By exploiting them directly, grouping neurons through shared hyperedges, communication traffic and hardware resource usage can be reduced be yond what just contracting individual connections attains. To substantiate this insight, we consider several partitioning and placement algorithms, some newly devised, others adapted from literature, and compare them over progressively larger and bio-plausible SNNs. Our results show that hypergraph based techniques can achieve better mappings than the state-of-the-art at several execution time regimes. Based on these observations, we identify a promising selection of algorithms to achieve effective mappings at any scale.

Marco Ronzani, Cristina Silvano

Hypergraph partitioning is a pervasive NP-hard problem, and accelerating its computation on GPU can both slice time-to-solution and raise quality of results. In this work, we implement a multi-level hypergraph partitioning algorithm on GPU targeting a specific set of problem constraints: bounded per-partition size and distinct inbound hyperedges. Manipulating hypergraphs requires long orders of nested iterations, and enforcing these constraints introduces further set operations amidst them. Hence, we design algorithms around our problem's specifics, materializing the hypergraph's incidence structure in memory and exploiting set sparsity. Our results show competitive speedups as high as 940x and 2-26% better results in connectivity over a sequential multi-level partitioner.

Amir H. Ashouri, William Killian, John Cavazos et al.

Since the mid-1990s, researchers have been trying to use machine-learning based approaches to solve a number of different compiler optimization problems. These techniques primarily enhance the quality of the obtained results and, more importantly, make it feasible to tackle two main compiler optimization problems: optimization selection (choosing which optimizations to apply) and phase-ordering (choosing the order of applying optimizations). The compiler optimization space continues to grow due to the advancement of applications, increasing number of compiler optimizations, and new target architectures. Generic optimization passes in compilers cannot fully leverage newly introduced optimizations and, therefore, cannot keep up with the pace of increasing options. This survey summarizes and classifies the recent advances in using machine learning for the compiler optimization field, particularly on the two major problems of (1) selecting the best optimizations and (2) the phase-ordering of optimizations. The survey highlights the approaches taken so far, the obtained results, the fine-grain classification among different approaches and finally, the influential papers of the field.