اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدBalancing Efficiency and Flexibility for DNN Acceleration via Temporal GPU-Systolic Array Integration
59
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
The research interest in specialized hardware accelerators for deep neural networks (DNN) spikes recently owing to their superior performance and efficiency. However, todays DNN accelerators primarily focus on accelerating specific kernels such as convolution and matrix multiplication, which are vital but only part of an end-to-end DNN-enabled application. Meaningful speedups over the entire application often require supporting computations that are, while massively parallel, ill-suited to DNN accelerators. Integrating a general-purpose processor such as a CPU or a GPU incurs significant data movement overhead and leads to resource under-utilization on the DNN accelerators. We propose Simultaneous Multi-mode Architecture (SMA), a novel architecture design and execution model that offers general-purpose programmability on DNN accelerators in order to accelerate end-to-end applications. The key to SMA is the temporal integration of the systolic execution model with the GPU-like SIMD execution model. The SMA exploits the common components shared between the systolic-array accelerator and the GPU, and provides lightweight reconfiguration capability to switch between the two modes in-situ. The SMA achieves up to 63% performance improvement while consuming 23% less energy than the baseline Volta architecture with TensorCore.
قيم البحث
اقرأ أيضاً
Convolutional neural network (CNN) inference on mobile devices demands efficient hardware acceleration of low-precision (INT8) general matrix multiplication (GEMM). Exploiting data sparsity is a common approach to further accelerate GEMM for CNN infe
rence, and in particular, structural sparsity has the advantages of predictable load balancing and very low index overhead. In this paper, we address a key architectural challenge with structural sparsity: how to provide support for a range of sparsity levels while maintaining high utilization of the hardware. We describe a time unrolled formulation of variable density-bound block (VDBB) sparsity that allows for a configurable number of non-zero elements per block, at constant utilization. We then describe a systolic array microarchitecture that implements this scheme, with two data reuse optimizations. Firstly, we increase reuse in both operands and partial products by increasing the number of MACs per PE. Secondly, we introduce a novel approach of moving the IM2COL transform into the hardware, which allows us to achieve a 3x data bandwidth expansion just before the operands are consumed by the datapath, reducing the SRAM power consumption. The optimizations for weight sparsity, activation sparsity and data reuse are all interrelated and therefore the optimal combination is not obvious. Therefore, we perform an design space evaluation to find the pareto-optimal design characteristics. The resulting design achieves 16.8 TOPS/W in 16nm with modest 50% model sparsity and scales with model sparsity up to 55.7TOPS/W at 87.5%. As well as successfully demonstrating the variable DBB technique, this result significantly outperforms previously reported sparse CNN accelerators.
In this paper, we use graphics processing units(GPU) to accelerate sparse and arbitrary structured neural networks. Sparse networks have nodes in the network that are not fully connected with nodes in preceding and following layers, and arbitrary str
ucture neural networks have different number of nodes in each layers. Sparse Neural networks with arbitrary structures are generally created in the processes like neural network pruning and evolutionary machine learning strategies. We show that we can gain significant speedup for full activation of such neural networks using graphical processing units. We do a prepossessing step to determine dependency groups for all the nodes in a network, and use that information to guide the progression of activation in the neural network. Then we compute activation for each nodes in its own separate thread in the GPU, which allows for massive parallelization. We use CUDA framework to implement our approach and compare the results of sequential and GPU implementations. Our results show that the activation of sparse neural networks lends very well to GPU acceleration and can help speed up machine learning strategies which generate such networks or other processes that have similar structure.
We demonstrate the first implementation of recently-developed fast explicit kinetic integration algorithms on modern graphics processing unit (GPU) accelerators. Taking as a generic test case a Type Ia supernova explosion with an extremely stiff ther
monuclear network having 150 isotopic species and 1604 reactions coupled to hydrodynamics using operator splitting, we demonstrate the capability to solve of order 100 realistic kinetic networks in parallel in the same time that standard implicit methods can solve a single such network on a CPU. This orders-of-magnitude decrease in compute time for solving systems of realistic kinetic networks implies that important coupled, multiphysics problems in various scientific and technical fields that were intractible, or could be simulated only with highly schematic kinetic networks, are now computationally feasible.
With widespread advances in machine learning, a number of large enterprises are beginning to incorporate machine learning models across a number of products. These models are typically trained on shared, multi-tenant GPU clusters. Similar to existing
cluster computing workloads, scheduling frameworks aim to provide features like high efficiency, resource isolation, fair sharing across users, etc. However Deep Neural Network (DNN) based workloads, predominantly trained on GPUs, differ in two significant ways from traditional big data analytics workloads. First, from a cluster utilization perspective, GPUs represent a monolithic resource that cannot be shared at a fine granularity across users. Second, from a workload perspective, deep learning frameworks require gang scheduling reducing the flexibility of scheduling and making the jobs themselves inelastic to failures at runtime. In this paper we present a detailed workload characterization of a two-month long trace from a multi-tenant GPU cluster in a large enterprise. By correlating scheduler logs with logs from individual jobs, we study three distinct issues that affect cluster utilization for DNN training workloads on multi-tenant clusters: (1) the effect of gang scheduling and locality constraints on queuing, (2) the effect of locality on GPU utilization, and (3) failures during training. Based on our experience running a large-scale operation, we provide design guidelines pertaining to next-generation cluster schedulers for DNN training workloads.
Deep neural networks (DNNs) sustain high performance in todays data processing applications. DNN inference is resource-intensive thus is difficult to fit into a mobile device. An alternative is to offload the DNN inference to a cloud server. However,
such an approach requires heavy raw data transmission between the mobile device and the cloud server, which is not suitable for mission-critical and privacy-sensitive applications such as autopilot. To solve this problem, recent advances unleash DNN services using the edge computing paradigm. The existing approaches split a DNN into two parts and deploy the two partitions to computation nodes at two edge computing tiers. Nonetheless, these methods overlook collaborative device-edge-cloud computation resources. Besides, previous algorithms demand the whole DNN re-partitioning to adapt to computation resource changes and network dynamics. Moreover, for resource-demanding convolutional layers, prior works do not give a parallel processing strategy without loss of accuracy at the edge side. To tackle these issues, we propose D3, a dynamic DNN decomposition system for synergistic inference without precision loss. The proposed system introduces a heuristic algorithm named horizontal partition algorithm to split a DNN into three parts. The algorithm can partially adjust the partitions at run time according to processing time and network conditions. At the edge side, a vertical separation module separates feature maps into tiles that can be independently run on different edge nodes in parallel. Extensive quantitative evaluation of five popular DNNs illustrates that D3 outperforms the state-of-the-art counterparts up to 3.4 times in end-to-end DNN inference time and reduces backbone network communication overhead up to 3.68 times.
الأسئلة المقترحة
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
