The record-breaking performance of deep neural networks (DNNs) comes with heavy parameter budgets, which leads to external dynamic random access memory (DRAM) for storage. The prohibitive energy of DRAM accesses makes it nontrivial for DNN deployment on resource-constrained devices, calling for minimizing the movements of weights and data in order to improve the energy efficiency. Driven by this critical bottleneck, we present SmartDeal, a hardware-friendly algorithm framework to trade higher-cost memory storage/access for lower-cost computation, in order to aggressively boost the storage and energy efficiency, for both DNN inference and training. The core technique of SmartDeal is a novel DNN weight matrix decomposition framework with respective structural constraints on each matrix factor, carefully crafted to unleash the hardware-aware efficiency potential. Specifically, we decompose each weight tensor as the product of a small basis matrix and a large structurally sparse coefficient matrix whose nonzero elements are readily quantized to the power-of-2. The resulting sparse and readily quantized DNNs enjoy greatly reduced energy consumption in data movement as well as weight storage, while incurring minimal overhead to recover the original weights thanks to the required sparse bit-operations and cost-favorable computations. Beyond inference, we take another leap to embrace energy-efficient training, by introducing several customized techniques to address the unique roadblocks arising in training while preserving the SmartDeal structures. We also design a dedicated hardware accelerator to fully utilize the new weight structure to improve the real energy efficiency and latency performance. We conduct experiments on both vision and language tasks, with nine models, four datasets, and three settings (inference-only, adaptation, and fine-tuning). Our extensive results show that 1) being applied to inference, SmartDeal achieves up to 2.44x improvement in energy efficiency as evaluated using real hardware implementations and 2) being applied to training, SmartDeal can lead to 10.56x and 4.48x reduction in the storage and the training energy cost, respectively, with usually negligible accuracy loss, compared to state-of-the-art training baselines. Our source codes are available at: https://github.com/VITA-Group/SmartDeal.
more »
« less
This content will become publicly available on December 31, 2023
ASA: A ccelerating S parse A ccumulation in Column-wise SpGEMM
Sparse linear algebra is an important kernel in many different applications. Among various sparse general matrix-matrix multiplication (SpGEMM) algorithms, Gustavson’s column-wise SpGEMM has good locality when reading input matrix and can be easily parallelized by distributing the computation of different columns of an output matrix to different processors. However, the sparse accumulation (SPA) step in column-wise SpGEMM, which merges partial sums from each of the multiplications by the row indices, is still a performance bottleneck. The state-of-the-art software implementation uses a hash table for partial sum search in the SPA, which makes SPA the largest contributor to the execution time of SpGEMM. There are three reasons that cause the SPA to become the bottleneck: (1) hash probing requires data-dependent branches that are difficult for a branch predictor to predict correctly; (2) the accumulation of partial sum is dependent on the results of the hash probing, which makes it difficult to hide the hash probing latency; and (3) hash collision requires time-consuming linear search and optimizations to reduce these collisions require an accurate estimation of the number of non-zeros in each column of the output matrix. This work proposes ASA architecture to accelerate the SPA. ASA overcomes the challenges of SPA by (1) executing the partial sum search and accumulate with a single instruction through ISA extension to eliminate data-dependent branches in hash probing, (2) using a dedicated on-chip cache to perform the search and accumulation in a pipelined fashion, (3) relying on the parallel search capability of a set-associative cache to reduce search latency, and (4) delaying the merging of overflowed entries. As a result, ASA achieves an average of 2.25× and 5.05× speedup as compared to the state-of-the-art software implementation of a Markov clustering application and its SpGEMM kernel, respectively. As compared to a state-of-the-art hashing accelerator design, ASA achieves an average of 1.95× speedup in the SpGEMM kernel.
more »
« less
- Award ID(s):
- 1750826
- NSF-PAR ID:
- 10407645
- Date Published:
- Journal Name:
- ACM Transactions on Architecture and Code Optimization
- Volume:
- 19
- Issue:
- 4
- ISSN:
- 1544-3566
- Page Range / eLocation ID:
- 1 to 24
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The continuous growth of CNN complexity not only intensifies the need for hardware acceleration but also presents a huge challenge. That is, the solution space for CNN hardware design and dataflow mapping becomes enormously large besides the fact that it is discrete and lacks a well behaved structure. Most previous works either are stochastic metaheuristics, such as genetic algorithm, which are typically very slow for solving large problems, or rely on expensive sampling, e.g., Gumbel Softmax-based differentiable optimization and Bayesian optimization. We propose an analytical model for evaluating power and performance of CNN hardware design and dataflow solutions. Based on this model, we introduce a co-optimization method consisting of nonlinear programming and parallel local search. A key innovation in this model is its matrix form, which enables the use of deep learning toolkit for highly efficient computations of power/performance values and gradients in the optimization. In handling power-performance tradeoff, our method can lead to better solutions than minimizing a weighted sum of power and latency. The average relative error of our model compared with Timeloop is as small as 1%. Compared to state-of-the-art methods, our approach achieves solutions with up to 1.7 × shorter inference latency, 37.5% less power consumption, and 3 × less area on ResNet 18. Moreover, it provides a 6.2 × speedup of optimizationmore » « less
-
Breath-first search (BFS) is a fundamental building block in many graph-based applications. It is challenging to optimize due to its irregular memory-access pattern. Prior work, based on hardware description languages (HDLs) and high-level synthesis (HLS), address the memory-access bottleneck by using techniques such as edge-centric traversal, data alignment, and compute-unit (CU) replication. While these optimizations work well for dense graph datasets, optimizing BFS on sparse graphs remains a significant challenge due to the kernel launch overhead and poor workload distribution across processing elements. As a complement to the prior work, we present and evaluate optimizations in OpenCL for BFS on sparse graphs. Specifically, we explore application-specific and architecture-aware optimizations aimed at mitigating the irregular global-memory access bottleneck in sparse graphs. In our kernel design, we consider factors such as choice of data structure between queue and array, number of memory banks, and kernel launch configuration. We evaluate the impact of proposed optimizations on a diverse set of sparse graphs. In comparison with the state-of-the-art OpenCL implementation for FPGA, we achieve 5.7x-22.3x speedup on Stratix 10 SX 2800 FPGA for the graphs that are most sensitive to our optimization scheme.more » « less
-
Neural architecture search (NAS) is a promising technique to design efficient and high-performance deep neural networks (DNNs). As the performance requirements of ML applications grow continuously, the hardware accelerators start playing a central role in DNN design. This trend makes NAS even more complicated and time-consuming for most real applications. This paper proposes FLASH, a very fast NAS methodology that co-optimizes the DNN accuracy and performance on a real hardware platform. As the main theoretical contribution, we first propose the NN-Degree, an analytical metric to quantify the topological characteristics of DNNs with skip connections (e.g., DenseNets, ResNets, Wide-ResNets, and MobileNets). The newly proposed NN-Degree allows us to do training-free NAS within one second and build an accuracy predictor by training as few as 25 samples out of a vast search space with more than 63 billion configurations. Second, by performing inference on the target hardware, we fine-tune and validate our analytical models to estimate the latency, area, and energy consumption of various DNN architectures while executing standard ML datasets. Third, we construct a hierarchical algorithm based on simplicial homology global optimization (SHGO) to optimize the model-architecture co-design process, while considering the area, latency, and energy consumption of the target hardware. We demonstrate that, compared to the state-of-the-art NAS approaches, our proposed hierarchical SHGO-based algorithm enables more than four orders of magnitude speedup (specifically, the execution time of the proposed algorithm is about 0.1 seconds). Finally, our experimental evaluations show that FLASH is easily transferable to different hardware architectures, thus enabling us to do NAS on a Raspberry Pi-3B processor in less than 3 seconds.more » « less
-
Vector search has drawn a rapid increase of interest in the research community due to its application in novel AI applications. Maximizing its performance is essential for many tasks but remains preliminary understood. In this work, we investigate the root causes of the scalability bottleneck of using intra-query parallelism to speedup the state-of-the-art graph-based vector search systems on multi-core architectures. Our in-depth analysis reveals several scalability challenges from both system and algorithm perspectives. Based on the insights, we propose iQAN, a parallel search algorithm with a set of optimizations that boost convergence, avoid redundant computations, and mitigate synchronization overhead. Our evaluation results on a wide range of real-world datasets show that iQAN achieves up to 37.7× and 76.6× lower latency than state-of-the-art sequential baselines on datasets ranging from a million to a hundred million datasets. We also show that iQAN achieves outstanding scalability as the graph size or the accuracy target increases, allowing it to outperform the state-of-the-art baseline on two billion-scale datasets by up to 16.0× with up to 64 cores.more » « less