Search

Neural networks (NNs) are growing in importance and complexity. A neural network's performance (and energy efficiency) can be bound either by computation or memory resources. The processing-in-memory (PIM) paradigm, where computation is placed near or within memory arrays, is a viable solution to accelerate memory-bound NNs. However, PIM architectures vary in form, where different PIM approaches lead to different trade-offs. Our goal is to analyze, discuss, and contrast DRAM-based PIM architectures for NN performance and energy efficiency. To do so, we analyze three state-of-the-art PIM architectures: (1) UPMEM, which integrates processors and DRAM arrays into a single 2D chip; (2) Mensa, a 3D-stack-based PIM architecture tailored for edge devices; and (3) SIMDRAM, which uses the analog principles of DRAM to execute bit-serial operations. Our analysis reveals that PIM greatly benefits memory-bound NNs: (1) UPMEM provides 23x the performance of a high-end GPU when the GPU requires memory oversubscription for a general matrix-vector multiplication kernel; (2) Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Google Edge TPU for 24 Google edge NN models; and (3) SIMDRAM outperforms a CPU/GPU by 16.7x/1.4x for three binary NNs. We conclude that the ideal PIM architecture for NN models depends on a model's distinct attributes, due to the inherent architectural design choices., This is an extended and updated version of a paper published in IEEE Micro, pp. 1-14, 29 Aug. 2022. arXiv admin note: text overlap with arXiv:2109.14320

Partitioning applications between NDP and host CPU cores causes inter-segment data movement overhead, which is caused by moving data generated from one segment (e.g., instructions, functions) and used in consecutive segments. Prior works take two approaches to this problem. The first class of works maps segments to NDP or host cores based on the properties of each segment, neglecting the inter-segment data movement overhead. The second class of works partitions applications based on the overall memory bandwidth saving of each segment, and does not offload each segment to the best-fitting core if they incur high inter-segment data movement. We show that 1) mapping each segment to its best-fitting core ideally can provide substantial benefits, and 2) the inter-segment data movement reduces this benefit significantly. To this end, we introduce ALP, a new programmer-transparent technique to leverage the performance benefits of NDP by alleviating the inter-segment data movement overhead between host and memory and enabling efficient partitioning of applications. ALP alleviates the inter-segment data movement overhead by proactively and accurately transferring the required data between the segments. This is based on the key observation that the instructions that generate the inter-segment data stay the same across different executions of a program on different inputs. ALP uses a compiler pass to identify these instructions and uses specialized hardware to transfer data between the host and NDP cores at runtime. ALP efficiently maps application segments to either host or NDP considering 1) the properties of each segment, 2) the inter-segment data movement overhead, and 3) whether this overhead can be alleviated in a timely manner. We evaluate ALP across a wide range of workloads and show on average 54.3% and 45.4% speedup compared to only-host CPU or only-NDP executions, respectively., To appear in IEEE TETC

Graphics Processing Units (GPUs) are widely-used accelerators for data-parallel applications. In many GPU applications, GPU memory bandwidth bottlenecks performance, causing underutilization of GPU cores. Hence, disabling many cores does not affect the performance of memory-bound workloads. While simply power-gating unused GPU cores would save energy, prior works attempt to better utilize GPU cores for other applications (ideally compute-bound), which increases the GPU's total throughput. In this paper, we introduce Morpheus, a new hardware/software co-designed technique to boost the performance of memory-bound applications. The key idea of Morpheus is to exploit unused core resources to extend the GPU last level cache (LLC) capacity. In Morpheus, each GPU core has two execution modes: compute mode and cache mode. Cores in compute mode operate conventionally and run application threads. However, for the cores in cache mode, Morpheus invokes a software helper kernel that uses the cores' on-chip memories (i.e., register file, shared memory, and L1) in a way that extends the LLC capacity for a running memory-bound workload. Morpheus adds a controller to the GPU hardware to forward LLC requests to either the conventional LLC (managed by hardware) or the extended LLC (managed by the helper kernel). Our experimental results show that Morpheus improves the performance and energy efficiency of a baseline GPU architecture by an average of 39% and 58%, respectively, across several memory-bound workloads. Morpheus' performance is within 3% of a GPU design that has a quadruple-sized conventional LLC. Morpheus can thus contribute to reducing the hardware dedicated to a conventional LLC by exploiting idle cores' on-chip memory resources as additional cache capacity., Comment: To appear in 55th IEEE/ACM International Symposium on Microarchitecture (MICRO), 2022

Bulk bitwise operations, i.e., bitwise operations on large bit vectors, are prevalent in a wide range of important application domains, including databases, graph processing, genome analysis, cryptography, and hyper-dimensional computing. In conventional systems, the performance and energy efficiency of bulk bitwise operations are bottlenecked by data movement between the compute units and the memory hierarchy. In-flash processing (i.e., processing data inside NAND flash chips) has a high potential to accelerate bulk bitwise operations by fundamentally reducing data movement through the entire memory hierarchy. We identify two key limitations of the state-of-the-art in-flash processing technique for bulk bitwise operations; (i) it falls short of maximally exploiting the bit-level parallelism of bulk bitwise operations; (ii) it is unreliable because it does not consider the highly error-prone nature of NAND flash memory. We propose Flash-Cosmos (Flash Computation with One-Shot Multi-Operand Sensing), a new in-flash processing technique that significantly increases the performance and energy efficiency of bulk bitwise operations while providing high reliability. Flash-Cosmos introduces two key mechanisms that can be easily supported in modern NAND flash chips: (i) Multi-Wordline Sensing (MWS), which enables bulk bitwise operations on a large number of operands with a single sensing operation, and (ii) Enhanced SLC-mode Programming (ESP), which enables reliable computation inside NAND flash memory. We demonstrate the feasibility of performing bulk bitwise operations with high reliability in Flash-Cosmos by testing 160 real 3D NAND flash chips. Our evaluation shows that Flash-Cosmos improves average performance and energy efficiency by 3.5x/32x and 3.3x/95x, respectively, over the state-of-the-art in-flash/outside-storage processing techniques across three real-world applications., Comment: To appear in 55th IEEE/ACM International Symposium on Microarchitecture (MICRO), 2022

The increasing prevalence and growing size of data in modern applications have led to high costs for computation in traditional processor-centric computing systems. Moving large volumes of data between memory devices (e.g., DRAM) and computing elements (e.g., CPUs, GPUs) across bandwidth-limited memory channels can consume more than 60% of the total energy in modern systems. To mitigate these costs, the processing-in-memory (PIM) paradigm moves computation closer to where the data resides, reducing (and in some cases eliminating) the need to move data between memory and the processor. There are two main approaches to PIM: (1) processing-near-memory (PnM), where PIM logic is added to the same die as memory or to the logic layer of 3D-stacked memory; and (2) processing-using-memory (PuM), which uses the operational principles of memory cells to perform computation. Many works from academia and industry have shown the benefits of PnM and PuM for a wide range of workloads from different domains. However, fully adopting PIM in commercial systems is still very challenging due to the lack of tools and system support for PIM architectures across the computer architecture stack, which includes: (i) workload characterization methodologies and benchmark suites targeting PIM architectures; (ii) frameworks that can facilitate the implementation of complex operations and algorithms using the underlying PIM primitives; (iii) compiler support and compiler optimizations targeting PIM architectures; (iv) operating system support for PIM-aware virtual memory, memory management, data allocation, and data mapping; and (v) efficient data coherence and consistency mechanisms. Our goal in this work is to provide tools and system support for PnM and PuM architectures, aiming to ease the adoption of PIM in current and future systems., Comment: arXiv admin note: substantial text overlap with arXiv:2012.11890

Time series analysis is a key technique for extracting and predicting events in domains as diverse as epidemiology, genomics, neuroscience, environmental sciences, economics, and more. Matrix profile, the state-of-the-art algorithm to perform time series analysis, computes the most similar subsequence for a given query subsequence within a sliced time series. Matrix profile has low arithmetic intensity, but it typically operates on large amounts of time series data. In current computing systems, this data needs to be moved between the off-chip memory units and the on-chip computation units for performing matrix profile. This causes a major performance bottleneck as data movement is extremely costly in terms of both execution time and energy. In this work, we present NATSA, the first Near-Data Processing accelerator for time series analysis. The key idea is to exploit modern 3D-stacked High Bandwidth Memory (HBM) to enable efficient and fast specialized matrix profile computation near memory, where time series data resides. NATSA provides three key benefits: 1) quickly computing the matrix profile for a wide range of applications by building specialized energy-efficient floating-point arithmetic processing units close to HBM, 2) improving the energy efficiency and execution time by reducing the need for data movement over slow and energy-hungry buses between the computation units and the memory units, and 3) analyzing time series data at scale by exploiting low-latency, high-bandwidth, and energy-efficient memory access provided by HBM. Our experimental evaluation shows that NATSA improves performance by up to 14.2x (9.9x on average) and reduces energy by up to 27.2x (19.4x on average), over the state-of-the-art multi-core implementation. NATSA also improves performance by 6.3x and reduces energy by 10.2x over a general-purpose NDP platform with 64 in-order cores., Comment: To appear in ISVLSI 2022 Special Session on Processing in Memory. arXiv admin note: text overlap with arXiv:2010.02079

Many modern workloads such as neural network inference and graph processing are fundamentally memory-bound. For such workloads, data movement between memory and CPU cores imposes a significant overhead in terms of both latency and energy. A major reason is that this communication happens through a narrow bus with high latency and limited bandwidth, and the low data reuse in memory-bound workloads is insufficient to amortize the cost of memory access. Fundamentally addressing this data movement bottleneck requires a paradigm where the memory system assumes an active role in computing by integrating processing capabilities. This paradigm is known as processing-in-memory (PIM). Recent research explores different forms of PIM architectures, motivated by the emergence of new technologies that integrate memory with a logic layer, where processing elements can be easily placed. Past works evaluate these architectures in simulation or, at best, with simplified hardware prototypes. In contrast, the UPMEM company has designed and manufactured the first publicly-available real-world PIM architecture. The UPMEM PIM architecture combines traditional DRAM memory arrays with general-purpose in-order cores, called DRAM Processing Units (DPUs), integrated in the same chip. This paper presents key takeaways from the first comprehensive analysis of the first publicly-available real-world PIM architecture. We provide four key takeaways about the UPMEM PIM architecture, which stem from our study. More insights about suitability of different workloads to the PIM system, programming recommendations for software designers, and suggestions and hints for hardware and architecture designers of future PIM systems are available in arXiv:2105.03814, Comment: Invited paper to appear at Workshop on Computing with Unconventional Technologies (CUT) 2021 https://sites.google.com/umn.edu/cut-2021/home. arXiv admin note: substantial text overlap with arXiv:2105.03814

In-place data manipulation is very desirable in many-core architectures with limited on-board memory. This paper deals with the in-place implementation of a class of primitives that perform data movements in one direction. We call these primitives Data Sliding (DS) algorithms. Notable among them are relational algebra primitives (such as select and unique), padding to insert empty elements in a data structure, and stream compaction to reduce memory requirements. Their in-place implementation in a bulk synchronous parallel model, such as GPUs, is specially challenging due to the difficulties in synchronizing threads executing on different compute units. Using a novel adjacent work-group synchronization technique, we propose two algorithmic schemes for regular and irregular DS algorithms. With a set of 5 benchmarks, we validate our approaches and compare them to the state-of-the-art implementations of these benchmarks. Our regular DS algorithms demonstrate up to 9.11x and 73.25x on NVIDIA and AMD GPUs, respectively, the throughput of their competitors. Our irregular DS algorithms outperform NVIDIA Thrust library by up to 3.24x on the three most recent generations of NVIDIA GPUs.