Research has shown that convolutional neural networks contain significant redundancy, and high classification accuracy can be obtained even when weights and activations are reduced from floating point to binary values. In this paper, we present Finn, a framework for building fast and flexible FPGA accelerators using a flexible heterogeneous streaming architecture. By utilizing a novel set of optimizations that enable efficient mapping of binarized neural networks to hardware, we implement fully connected, convolutional and pooling layers, with per-layer compute resources being tailored to user-provided throughput requirements. On a ZC706 embedded FPGA platform drawing less than 25 W total system power, we demonstrate up to 12.3 million image classifications per second with 0.31 µs latency on the MNIST dataset with 95.8% accuracy, and 21906 image classifications per second with 283 µs latency on the CIFAR-10 and SVHN datasets with respectively 80.1% and 94.9% accuracy. To the best of our knowledge, ours are the fastest classification rates reported to date on these benchmarks.

Binarized neural networks (BNNs) are gaining interest in the deep learning community due to their significantly lower computational and memory cost. They are particularly well suited to reconfigurable logic devices, which contain an abundance of fine-grained compute resources and can result in smaller, lower power implementations, or conversely in higher classification rates. Towards this end, the Finn framework was recently proposed for building fast and flexible field programmable gate array (FPGA) accelerators for BNNs. Finn utilized a novel set of optimizations that enable efficient mapping of BNNs to hardware and implemented fully connected, non-padded convolutional and pooling layers, with per-layer compute resources being tailored to user-provided throughput requirements. However, FINN was not evaluated on larger topologies due to the size of the chosen FPGA, and exhibited decreased accuracy due to lack of padding. In this paper, we improve upon Finn to show how padding can be employed on BNNs while still maintaining a 1-bit datapath and high accuracy. Based on this technique, we demonstrate numerous experiments to illustrate flexibility and scalability of the approach. In particular, we show that a large BNN requiring 1.2 billion operations per frame running on an ADM-PCIE-8K5 platform can classify images at 12 kFPS with 671 µs latency while drawing less than 41 W board power and classifying CIFAR-10 images at 88.7% accuracy. Our implementation of this network achieves 14.8 trillion operations per second. We believe this is the fastest classification rate reported to date on this benchmark at this level of accuracy.

No abstract

Energy efficiency is an important aspect of future exascale systems, mainly due to rising energy cost. Although High performance computing (HPC) applications are compute centric, they still exhibit varying computational characteristics in different regions of the program, such as compute-, memory-, and I/O-bound code regions. Some of today's clusters already offer mechanisms to adjust the system to the resource requirements of an application, e.g., by controlling the CPU frequency. However, manually tuning for improved energy efficiency is a tedious and painstaking task that is often neglected by application developers. The European Union's Horizon 2020 project READEX (Runtime Exploitation of Application Dynamism for Energyefficient eXascale computing) aims at developing a tools-aided approach for improved energy efficiency of current and future HPC applications. To reach this goal, the READEX project combines technologies from two ends of the compute spectrum, embedded systems and HPC, constituting a split design-time/runtime methodology. dynamic auto-tuning of fine-grained application regions using the systems scenario methodology, which was originally developed for improving the energy efficiency in embedded systems. This paper introduces the concepts of the READEX project, its envisioned implementation, and preliminary results that demonstrate the feasibility of this approach.

GPU memory systems adopt a multi-dimensional hardware structure to provide the bandwidth necessary to support 100s to 1000s of concurrent threads. On the software side, GPU-compute workloads also use multi-dimensional structures to organize the threads. We observe that these structures can combine unfavorably and create significant resource imbalance in the memory subsystem-causing low performance and poor power-efficiency. The key issue is that it is highly applicationdependent which memory address bits exhibit high variability. To solve this problem, we first provide an entropy analysis approach tailored for the highly concurrent memory request behavior in GPU-compute workloads. Our window-based entropy metric captures the information content of each address bit of the memory requests that are likely to co-exist in the memory system at runtime. Using this metric, we find that GPU-compute workloads exhibit entropy valleys distributed throughout the lower order address bits. This indicates that efficient GPU-address mapping schemes need to harvest entropy from broad address-bit ranges and concentrate the entropy into the bits used for channel and bank selection in the memory subsystem. This insight leads us to propose the Page Address Entropy (PAE) mapping scheme which concentrates the entropy of the row, channel and bank bits of the input address into the bank and channel bits of the output address. PAE maps straightforwardly to hardware and can be implemented with a tree of XOR-gates. PAE improves performance by 1.31× and power-efficiency by 1.25× compared to state-of-the-art permutation-based address mapping.

No abstract

No abstract

Abstract-Multi-core memory systems commonly share resources between processors. Resource sharing improves utilization at the cost of increased inter-application interference which may lead to priority inversion, missed deadlines and unpredictable interactive performance. A key component to effectively manage multi-core resources is performance accounting which aims to accurately estimate interference-free application performance. Previously proposed accounting systems are either invasive or transparent. Invasive accounting systems can be accurate, but slow down latency-sensitive processes. Transparent accounting systems do not affect performance, but tend to provide less accurate performance estimates.We propose a novel class of performance accounting systems that achieve both performance-transparency and superior accuracy. We call the approach dataflow accounting, and the key idea is to track dynamic dataflow properties and use these to estimate interference-free performance. Our main contribution is Graph-based Dynamic Performance (GDP) accounting. GDP dynamically builds a dataflow graph of load requests and periods where the processor commits instructions. This graph concisely represents the relationship between memory loads and forward progress in program execution. More specifically, GDP estimates interference-free stall cycles by multiplying the critical path length of the dataflow graph with the estimated interference-free memory latency. GDP is very accurate with mean IPC estimation errors of 3.4% and 9.8% for our 4-and 8-core processors, respectively. When GDP is used in a cache partitioning policy, we observe average system throughput improvements of 11.9% and 20.8% compared to partitioning using the state-of-the-art Application Slowdown Model.