This paper examines the design space trade-offs of DNNs accelerators aiming to achieve competitive performance and efficiency metrics for all four combinations of dense or sparse activation/weight tensors. To do so, we systematically examine the overheads of supporting sparsity on top of an optimized dense core. These overheads are modeled based on parameters that indicate how a multiplier can borrow a nonzero operation from the neighboring multipliers or future cycles. As a result of this exploration, we identify a few promising designs that perform better than prior work. Our findings suggest that even the best design targeting dual sparsity yields a 20%-30% drop in power efficiency when performing on single sparse models, i.e., those with only sparse weight or sparse activation tensors. We found that one can reuse resources of the same core to maintain high performance and efficiency when running single sparsity or dense models. We call this hybrid architecture Griffin. Griffin is 1.2, 3.0, 3.1, and 1.4x more power-efficient than state-of-the-art sparse architectures, for dense, weight-only sparse, activation-only sparse, and dual sparse models, respectively.

Industry is increasingly turning to reconfigurable architectures like FPGAs and CGRAs for improved performance and energy efficiency. Unfortunately, adoption of these architectures has been limited by their programming models. HDLs lack abstractions for productivity and are difficult to target from higher level languages. HLS tools are more productive, but offer an ad-hoc mix of software and hardware abstractions which make performance optimizations difficult.

Accelerating the inference of a trained DNN is a well studied subject. In this paper we switch the focus to the training of DNNs. The training phase is compute intensive, demands complicated data communication, and contains multiple levels of data dependencies and parallelism. This paper presents an algorithm/architecture space exploration of efficient accelerators to achieve better network convergence rates and higher energy efficiency for training DNNs. We further demonstrate that an architecture with hierarchical support for collective communication semantics provides flexibility in training various networks performing both stochastic and batched gradient descent based techniques. Our results suggest that smaller networks favor non-batched techniques while performance for larger networks is higher using batched operations. At 45nm technology, CATERPILLAR achieves performance efficiencies of 177 GFLOPS/W at over 80% utilization for SGD training on small networks and 211 GFLOPS/W at over 90% utilization for pipelined SGD/CP training on larger networks using a total area of 103.2 mm2and 178.9 mm2respectively.

Reconfigurable architectures have gained popularity in recent years as they allow the design of energy-efficient accelerators. Fine-grain fabrics (e.g. FPGAs) have traditionally suffered from performance and power inefficiencies due to bit-level reconfigurable abstractions. Both fine-grain and coarse-grain architectures (e.g. CGRAs) traditionally require low level programming and suffer from long compilation times. We address both challenges with Plasticine, a new spatially reconfigurable architecture designed to efficiently execute applications composed of parallel patterns. Parallel patterns have emerged from recent research on parallel programming as powerful, high-level abstractions that can elegantly capture data locality, memory access patterns, and parallelism across a wide range of dense and sparse applications.

State-of-the-art deep neural networks (DNNs) have hundreds of millions of connections and are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources and power budgets. While custom hardware helps the computation, fetching weights from DRAM is two orders of magnitude more expensive than ALU operations, and dominates the required power. Previously proposed 'Deep Compression' makes it possible to fit large DNNs (AlexNet and VGGNet) fully in on-chip SRAM. This compression is achieved by pruning the redundant connections and having multiple connections share the same weight. We propose an energy efficient inference engine (EIE) that performs inference on this compressed network model and accelerates the resulting sparse matrix-vector multiplication with weight sharing. Going from DRAM to SRAM gives EIE 120x energy saving, Exploiting sparsity saves 10x, Weight sharing gives 8x, Skipping zero activations from ReLU saves another 3x. Evaluated on nine DNN benchmarks, EIE is 189x and 13x faster when compared to CPU and GPU implementations of the same DNN without compression. EIE has a processing power of 102 GOPS working directly on a compressed network, corresponding to 3 TOPS on an uncompressed network, and processes FC layers of AlexNet at 1.88x104 frames/sec with a power dissipation of only 600mW. It is 24,000x and 3,400x more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9x, 19x and 3x better throughput, energy efficiency and area efficiency.

DRAM energy is an important component to optimize in modern computing systems. One outstanding source of DRAM energy is the energy to fetch data stored on cells to the row buffer, which occurs during two DRAM operations, row activate and refresh. This work exploits previously proposed half page row access, modifying the wordline connections within a bank to halve the number of cells fetched to the row buffer, to save energy in both cases. To accomplish this, we first change the data wire connections in the sub-array to reduce the cost of row buffer overfetch in multi-core systems which yields a 12% energy savings on average and a slight performance improvement in quad-core systems. We also propose charge recycling refresh, which reuses charge left over from a prior half page refresh to refresh another half page. Our charge recycling scheme is capable of reducing both auto- and self-refresh energy, saving more than 15% of refresh energy at 85°C, and provides even shorter refresh cycle time. Finally, we propose a refresh scheduling scheme that can dynamically adjust the number of charge recycled half pages, which can save up to 30% of refresh energy at 85°C.

Algorithms for computational imaging and computer vision are rapidly evolving, and hardware must follow suit: the next generation of image signal processors (ISPs) must be “programmable” to support new algorithms created with high-level frameworks. In this work, we compare flexible ISP architectures, using applications written in the Darkroom image processing language. We target two fundamental architecture classes: programmable in time, as represented by SIMD, and programmable in space, as typified by coarse grain reconfigurable array architectures (CGRA). We consider several optimizations on these two base architectures, such as register file partitioning for SIMD, bus based routing and pipelined wires for CGRA, and line buffer variations. After these optimizations on average, CGRA provides 1.6x better energy efficiency and 1.4x better compute density versus a SIMD solution, and 1.4x the energy efficiency and 3.1x the compute density of an FPGA. However the cost of providing general programmability is still high: compared to an ASIC, CGRA has 6x worse energy and area efficiency, and this ratio would be roughly 10x if memory dominated applications were excluded.

This paper examines the mapping of algorithms encountered when solving dense linear systems and linear least-squares problems to a custom Linear Algebra Processor. Specifically, the focus is on Cholesky, LU (with partial pivoting), and QR factorizations and their blocked algorithms. As part of the study, we expose the benefits of redesigning floating point units and their surrounding data-paths to support these complicated operations. We show how adding moderate complexity to the architecture greatly alleviates complexities in the algorithm. We study design tradeoffs and the effectiveness of architectural modifications to demonstrate that we can improve power and performance efficiency to a level that can otherwise only be expected of full-custom ASIC designs. A feasibility study of inner kernels is extended to blocked level and shows that, at block level, the Linear Algebra Core (LAC) can achieve high efficiencies with up to 45 GFLOPS/W for both Cholesky and LU factorization, and over 35 GFLOPS/W for QR factorization. While maintaining such efficiencies, our extensions to the MAC units can achieve up to 10, 12, and 20 percent speedup for the blocked algorithms of Cholesky, LU, and QR factorization, respectively.

This paper examines the mapping of algorithms encountered when solving dense linear systems and linear least-squares problems to a custom Linear Algebra Processor. Specifically, the focus is on Cholesky, LU (with partial pivoting), and QR factorizations. As part of the study, we expose the benefits of redesigning floating point units and their surrounding data-paths to support these complicated operations. We show how adding moderate complexity to the architecture greatly alleviates complexities in the algorithm. We study design trade-offs and the effectiveness of architectural modifications to demonstrate that we can improve power and performance efficiency to a level that can otherwise only be expected of full-custom ASIC designs. A feasibility study shows that our extensions to the MAC units can double the speed of required vector-norm operations while reducing energy by 60%. Similarly, up to 20% speedup with 15% savings in energy can be achieved for LU factorization. We show how such efficiency is maintained even in the complex inner kernels of these operations.

This paper considers the modifications required to transform a highly-efficient, specialized linear algebra core into an efficient engine for computing Fast Fourier Transforms (FFTs). We review the minimal changes required to support Radix-4 FFT computations and propose extensions to the micro-architecture of the baseline linear algebra core. Along the way, we study the critical differences between the two classes of algorithms. Special attention is paid to the configuration of the on-chip memory system to support high utilization. We examine design trade-offs between efficiency, specialization and flexibility, and their effects both on the core and memory hierarchy for a unified design as compared to dedicated accelerators for each application. The final design is a flexible architecture that can perform both classes of applications. Results show that the proposed hybrid FFT/Linear Algebra core can achieve 26.6 GFLOPS/S with a power efficiency of 40 GFLOPS/W, which is up to 100× and 40× more energy efficient than cutting-edge CPUs and GPUs, respectively.

Reducing power consumption and increasing efficiency is a key concern for many applications. It is well-accepted that specialization and heterogeneity are crucial strategies to improve both power and performance. Yet, how to design highly efficient processing elements while maintaining enough flexibility within a domain of applications is a fundamental question. In this paper, we present the design of a specialized Linear Algebra Core (LAC) for an important class of computational kernels, the level-3 Basic Linear Algebra Subprograms (BLAS). We demonstrate a detailed algorithm/architecture co-design for mapping a number of level-3 BLAS operations onto the LAC. Results show that our prototype LAC achieves a performance of around 64 GFLOPS (double precision) for these operations, while consuming less than 1.3 Watts in standard 45nm CMOS technology. This is on par with a full-custom design and up to 50× and 10× better in terms of power efficiency than CPUs and GPUs.

Reducing power consumption and increasing efficiency is a key concern for many applications. How to design highly efficient computing elements while maintaining enough flexibility within a domain of applications is a fundamental question. In this paper, we present how broadcast buses can eliminate the use of power hungry multi-ported register files in the context of data-parallel hardware accelerators for linear algebra operations. We demonstrate an algorithm/architecture co-design for the mapping of different collective communication operations, which are crucial for achieving performance and efficiency in most linear algebra routines, such as GEMM, SYRK and matrix transposition. We compare a broadcast bus based architecture with conventional SIMD, 2D-SIMD and flat register file for these operations in terms of area and energy efficiency. Results show that fast broadcast data movement abilities in a prototypical linear algebra core can achieve up to 75× better power and up to 10× better area efficiency compared to traditional SIMD architectures.

As technology is reaching physical limits, reducing power consumption is a key issue on our path to sustained performance. In this paper, we study fundamental tradeoffs and limits in efficiency (as measured in energy per operation) that can be achieved for an important class of kernels, namely the level-3 Basic Linear Algebra Subprograms (BLAS). It is well-accepted that specialization is the key to efficiency. This paper establishes a baseline by studying GEneral Matrix-matrix Multiplication (GEMM) on a variety of custom and general-purpose CPU and GPU architectures. Our analysis shows that orders of magnitude improvements in efficiency are possible with relatively simple customizations and fine-tuning of memory hierarchy configurations. We argue that these customizations can be generalized to perform other representative linear algebra operations. In addition to exposing the sources of inefficiencies in current CPUs and GPUs, our results show our prototype Linear Algebra Processor (LAP) implementing Double-precision GEMM (DGEMM) can achieve 600 GFLOPS while consuming less than 25 Watts in standard 45 nm technology, which is up to 50 × more energy efficient than cutting-edge CPUs.

Achieving high-performance while reducing power consumption is a key concern as technology scaling is reaching its limits. It is well-accepted that application-specific custom hardware can achieve orders of magnitude improvements in efficiency. The question is whether such efficiency can be maintained while providing enough flexibility to implement a broad class of operations. In this paper, we aim to answer this question for the domain of matrix computations. We propose a design of a novel linear algebra core and demonstrate that it can achieve orders of magnitude improvements in efficiency for matrix-matrix multiplication, an operation that is indicative for a broad class of matrix computations. A feasibility study shows that 47 double- and 104 single-precision GFLOPS/W can be achieved in 19.5 and 15.6 GFLOPS/mm2, respectively with current components and standard 45nm technology.