Search

We propose DFModel, a modeling framework for mapping dataflow computation graphs onto large-scale systems. Mapping a workload to a system requires optimizing dataflow mappings at various levels, including the inter-chip (between chips) level and the intra-chip (within a chip) level. DFModel is, to the best of our knowledge, the first framework to perform the optimization at multiple levels of the memory hierarchy and the interconnection network hierarchy. We use DFModel to explore a wide range of workloads on a variety of systems. Evaluated workloads include two state-of-the-art machine learning applications (Large Language Models and Deep Learning Recommendation Models) and two high-performance computing applications (High Performance LINPACK and Fast Fourier Transform). System parameters investigated span the combination of dataflow and traditional accelerator architectures, memory technologies (DDR, HBM), interconnect technologies (PCIe, NVLink), and interconnection network topologies (torus, DGX, dragonfly). For a variety of workloads on a wide range of systems, the DFModel provided a mapping that predicts an average of 1.25X better performance compared to the ones measured on real systems. DFModel shows that for large language model training, dataflow architectures achieve 1.52X higher performance, 1.59X better cost efficiency, and 1.6X better power efficiency compared to non-dataflow architectures. On an industrial system with dataflow architectures, the DFModel-optimized dataflow mapping achieves a speedup of 6.13X compared to non-dataflow mappings from previous performance models such as Calculon, and 1.52X compared to a vendor provided dataflow mapping.

EIE proposed to accelerate pruned and compressed neural networks, exploiting weight sparsity, activation sparsity, and 4-bit weight-sharing in neural network accelerators. Since published in ISCA'16, it opened a new design space to accelerate pruned and sparse neural networks and spawned many algorithm-hardware co-designs for model compression and acceleration, both in academia and commercial AI chips. In retrospect, we review the background of this project, summarize the pros and cons, and discuss new opportunities where pruning, sparsity, and low precision can accelerate emerging deep learning workloads., Comment: Invited retrospective paper at ISCA 2023

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., Comment: Accepted at the 28th IEEE International Symposium on High-Performance Computer Architecture - HPCA 2022

In this paper, we propose a mixed-precision convolution unit architecture which supports different integer and floating point (FP) precisions. The proposed architecture is based on low-bit inner product units and realizes higher precision based on temporal decomposition. We illustrate how to integrate FP computations on integer-based architecture and evaluate overheads incurred by FP arithmetic support. We argue that alignment and addition overhead for FP inner product can be significant since the maximum exponent difference could be up to 58 bits, which results into a large alignment logic. To address this issue, we illustrate empirically that no more than 26-bitproduct bits are required and up to 8-bit of alignment is sufficient in most inference cases. We present novel optimizations based on the above observations to reduce the FP arithmetic hardware overheads. Our empirical results, based on simulation and hardware implementation, show significant reduction in FP16 overhead. Over typical mixed precision implementation, the proposed architecture achieves area improvements of up to 25% in TFLOPS/mm2and up to 46% in TOPS/mm2with power efficiency improvements of up to 40% in TFLOPS/Wand up to 63% in TOPS/W., Comment: Accepted to appear in 4th Conference on Machine Learning and Systems 2021

This paper studies structured sparse training of CNNs with a gradual pruning technique that leads to fixed, sparse weight matrices after a set number of epochs. We simplify the structure of the enforced sparsity so that it reduces overhead caused by regularization. The proposed training methodology Campfire explores pruning at granularities within a convolutional kernel and filter. We study various tradeoffs with respect to pruning duration, level of sparsity, and learning rate configuration. We show that our method creates a sparse version of ResNet-50 and ResNet-50 v1.5 on full ImageNet while remaining within a negligible <1% margin of accuracy loss. To ensure that this type of sparse training does not harm the robustness of the network, we also demonstrate how the network behaves in the presence of adversarial attacks. Our results show that with 70% target sparsity, over 75% top-1 accuracy is achievable.

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 mm$^2$ and 178.9 mm$^2$ respectively., Comment: ASAP 2017: The 28th Annual IEEE International Conference on Application-specific Systems, Architectures and Processors

Convolutional Neural Networks (CNNs) are the state of the art solution for many computer vision problems, and many researchers have explored optimized implementations. Most implementations heuristically block the computation to deal with the large data sizes and high data reuse of CNNs. This paper explores how to block CNN computations for memory locality by creating an analytical model for CNN-like loop nests. Using this model we automatically derive optimized blockings for common networks that improve the energy efficiency of custom hardware implementations by up to an order of magnitude. Compared to traditional CNN CPU implementations based on highly-tuned, hand-optimized BLAS libraries,our x86 programs implementing the optimal blocking reduce the number of memory accesses by up to 90%.

The key challenge to improving performance in the age of Dark Silicon is how to leverage transistors when they cannot all be used at the same time. In modern SOCs, these transistors are often used to create specialized accelerators which improve energy efficiency for some applications by 10-1000X. While this might seem like the magic bullet we need, for most CPU applications more energy is dissipated in the memory system than in the processor: these large gains in efficiency are only possible if the DRAM and memory hierarchy are mostly idle. We refer to this desirable state as Dark Memory, and it only occurs for applications with an extreme form of locality. To show our findings, we introduce Pareto curves in the energy/op and mm$^2$/(ops/s) metric space for compute units, accelerators, and on-chip memory/interconnect. These Pareto curves allow us to solve the power, performance, area constrained optimization problem to determine which accelerators should be used, and how to set their design parameters to optimize the system. This analysis shows that memory accesses create a floor to the achievable energy-per-op. Thus high performance requires Dark Memory, which in turn requires co-design of the algorithm for parallelism and locality, with the hardware., Comment: 8 pages, To appear in IEEE Design and Test Journal

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 102GOPS/s working directly on a compressed network, corresponding to 3TOPS/s on an uncompressed network, and processes FC layers of AlexNet at 1.88x10^4 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., Comment: External Links: TheNextPlatform: http://goo.gl/f7qX0L ; O'Reilly: https://goo.gl/Id1HNT ; Hacker News: https://goo.gl/KM72SV ; Embedded-vision: http://goo.gl/joQNg8 ; Talk at NVIDIA GTC'16: http://goo.gl/6wJYvn ; Talk at Embedded Vision Summit: https://goo.gl/7abFNe ; Talk at Stanford University: https://goo.gl/6lwuer. Published as a conference paper in ISCA 2016