While neural network quantization effectively reduces the cost of matrix multiplications, aggressive quantization can expose non-matrix-multiply operations as significant performance and resource bottlenecks on embedded systems. Addressing such bottlenecks requires a comprehensive approach to tailoring the precision across operations in the inference computation. To this end, we introduce scaled-integer range analysis ( SIRA ), a static analysis technique employing interval arithmetic to determine the range, scale, and bias for tensors in quantized neural networks. We show how this information can be exploited to reduce the resource footprint of FPGA dataflow neural network accelerators via tailored bitwidth adaptation for accumulators and downstream operations, aggregation of scales and biases, and conversion of consecutive elementwise operations to thresholding operations. We integrate SIRA -driven optimizations into the open source FINN framework, then evaluate their effectiveness across a range of quantized neural network workloads and compare implementation alternatives for non-matrix-multiply operations. We demonstrate an average reduction of 17% for LUTs, 66% for DSPs, and 22% for accumulator bitwidths with SIRA optimizations, providing detailed benchmark analysis and analytical models to guide the implementation style for non-matrix layers. Finally, we open source SIRA to facilitate community exploration of its benefits across various applications and hardware platforms.

Recurrent neural networks (RNNs), particularly LSTMs, are effective for time-series tasks like sentiment analysis and short-term stock prediction. However, their computational complexity poses challenges for real-time deployment in resource constrained environments. While FPGAs offer a promising platform for energy-efficient AI acceleration, existing tools mainly target feed-forward networks, and LSTM acceleration typically requires full custom implementation. In this paper, we address this gap by leveraging the open-source and extensible FINN framework to enable the generalized deployment of LSTMs on FPGAs. Specifically, we leverage the Scan operator from the Open Neural Network Exchange (ONNX) specification to model the recurrent nature of LSTM computations, enabling support for mixed quantisation within them and functional verification of LSTM-based models. Furthermore, we introduce custom transformations within the FINN compiler to map the quantised ONNX computation graph to hardware blocks from the HLS kernel library of the FINN compiler and Vitis HLS. We validate the proposed tool-flow by training a quantised ConvLSTM model for a mid-price stock prediction task using the widely used dataset and generating a corresponding hardware IP of the model using our flow, targeting the XCZU7EV device. We show that the generated quantised ConvLSTM accelerator through our flow achieves a balance between performance (latency) and resource consumption, while matching (or bettering) inference accuracy of state-of-the-art models with reduced precision. We believe that the generalisable nature of the proposed flow will pave the way for resource-efficient RNN accelerator designs on FPGAs.

Abstract Most FPGA boards in the HPC domain are well-suited for parallel scaling because of the direct integration of versatile and high-throughput network ports. However, the utilization of their network capabilities is often challenging and error-prone because the whole network stack and communication patterns have to be implemented and managed on the FPGAs. Also, this approach conceptually involves a trade-off between the performance potential of improved communication and the impact of resource consumption for communication infrastructure, since the utilized resources on the FPGAs could otherwise be used for computations. In this work, we investigate this trade-off, firstly, by using synthetic benchmarks to evaluate the different configuration options of the communication framework ACCL and their impact on communication latency and throughput. Finally, we use our findings to implement a shallow water simulation whose scalability heavily depends on low-latency communication. With a suitable configuration of ACCL, good scaling behavior can be shown to all 48 FPGAs installed in the system. Overall, the results show that the availability of inter-FPGA communication frameworks as well as the configurability of framework and network stack are crucial to achieve the best application performance with low latency communication.

Post-training quantization (PTQ) is a powerful technique for model compression, reducing the numerical precision in neural networks without additional training overhead. Recent works have investigated adopting 8 -bit floating-point formats (FP8) in the context of PTQ for model inference. However, floating-point formats smaller than 8 bits and their relative comparison in terms of accuracy-hardware cost with integers remains unexplored on FPGAs. In this work, we present minifloats, which are reduced-precision floating-point formats capable of further reducing the memory footprint, latency, and energy cost of a model while approaching full-precision model accuracy. We implement a custom FPGA-based multiply-accumulate operator library and explore the vast design space, comparing minifloat and integer representations across 3 to 8 bits for both weights and activations. We also examine the applicability of various integer-based quantization techniques to minifloats. Our experiments show that minifloats offer a promising alternative for emerging workloads such as vision transformers.

Dilated and transposed convolutions are widely used in modern convolutional neural networks (CNNs). These kernels are used extensively during CNN training and inference of applications such as image segmentation and high-resolution image generation. We find that commonly-used low-power CNN inference accelerators arenotoptimized for both these convolutional kernels. Dilated and transposed convolutions introduce significant zero padding when mapped to the underlying spatial architecture, significantly degrading performance and energy efficiency. Existing approaches that address this issue require significant design changes to the otherwise simple, efficient, and well-adopted architectures used to compute direct convolutions. To address this challenge, we propose EcoFlow, a new set of dataflows and mapping algorithms for dilated and transposed convolutions. These algorithms are tailored to execute efficiently on existing low-cost, small-scale spatial architectures and requires minimal changes to existing accelerators. At its core, EcoFlow eliminates zero padding through careful dataflow orchestration and data mapping tailored to the spatial architecture. We evaluate EcoFlow on CNN training workloads and Generative Adversarial Network (GAN) workloads. Experiments in our new cycle-accurate simulator show that, using a common CNN inference accelerator, EcoFlow 1) reduces end-to-end CNN training time between 7-85%, and 2) improves end-to-end GAN training performance between 29-42%, compared to state-of-the-art CNN dataflows.

This work presents a novel reconfigurable architecture for Low Latency Graph Neural Network (LL-GNN) designs for particle detectors, delivering unprecedented low latency performance. Incorporating FPGA-based GNNs into particle detectors presents a unique challenge since it requires sub-microsecond latency to deploy the networks for online event selection with a data rate of hundreds of terabytes per second in the Level-1 triggers at the CERN Large Hadron Collider experiments. This article proposes a novel outer-product based matrix multiplication approach, which is enhanced by exploiting the structured adjacency matrix and a column-major data layout. In addition, we propose a custom code transformation for the matrix multiplication operations, which leverages the structured sparsity patterns and binary features of adjacency matrices to reduce latency and improve hardware efficiency. Moreover, a fusion step is introduced to further reduce the end-to-end design latency by eliminating unnecessary boundaries. Furthermore, a GNN-specific algorithm-hardware co-design approach is presented which not only finds a design with a much better latency but also finds a high accuracy design under given latency constraints. To facilitate this, a customizable template for this low latency GNN hardware architecture has been designed and open-sourced, which enables the generation of low-latency FPGA designs with efficient resource utilization using a high-level synthesis tool. Evaluation results show that our FPGA implementation is up to 9.0 times faster and achieves up to 13.1 times higher power efficiency than a GPU implementation. Compared to the previous FPGA implementations, this work achieves 6.51 to 16.7 times lower latency. Moreover, the latency of our FPGA design is sufficiently low to enable deployment of GNNs in a sub-microsecond, real-time collider trigger system, enabling it to benefit from improved accuracy. The proposed LL-GNN design advances the next generation of trigger systems by enabling sophisticated algorithms to process experimental data efficiently.

High-fan-in dot product computations are ubiquitous in highly relevant application domains, such as signal processing and machine learning. Particularly, the diverse set of data formats used in machine learning poses a challenge for flexible efficient design solutions. Ideally, a dot product summation is composed from a carry-free compressor tree followed by a terminal carry-propagate addition. On FPGA, these compressor trees are constructed from generalized parallel counters whose architecture is closely tied to the underlying reconfigurable fabric. This work reviews known counter designs and proposes new ones in the context of the new AMD Versal™ fabric. On this basis, we develop a compressor generator featuring variable-sized counters, novel counter composition heuristics, explicit clustering strategies, and case-specific optimizations like logic gate absorption. In comparison to the Vivado™ default implementation, the combination of such a compressor with a novel, highly efficient quaternary adder reduces the LUT footprint across different bit matrix input shapes by 45% for a plain summation and by 46% for a terminal accumulation at a slight cost in critical path delay still allowing an operation well above 500 MHz. We demonstrate the aptness of our solution at examples of low-precision integer dot product accumulation units.

A widely used approach to characterize input data in both databases and ML is computing the correlation between attributes. The operation is supported by all major database engines and ML platforms. However, it is an expensive operation as the number of attributes involved grows. To address the issue, in this paper we introduce AMNES, a stream analytics system offloading the correlation operator into an FPGA-based network interface card. AMNES processes data at network line rate and the design can be used in combination with smart storage or SmartNICs to implement near data or in-network data processing. AMNES design goes beyond matrix multiplication and offers a customized solution for correlation computation bypassing the CPU. Our experiments show that AMNES can sustain streams arriving at 100 Gbps over an RDMA network, while requiring only ten milliseconds to compute the correlation coefficients among 64 streams, an order of magnitude better than competing CPU or GPU designs.

Field-programmable gate array (FPGA)–based accelerators are becoming increasingly popular for deep neural network (DNN) inference due to their ability to scale performance with increasing degrees of specialization with dataflow architectures or custom data type precision. In order to reduce the barrier for software engineers and data scientists to adopt FPGAs, C++- and OpenCL-based design entries with high-level synthesis (HLS) have been introduced. They provide higher abstraction compared with register-transfer level (RTL)–based design. HLS offers faster development time, better maintainability, and more flexibility in code exploration when evaluating several options for multi-dimension tensors, convolutional layers, or different degrees of parallelism. For this reason, HLS has been adopted by DNN accelerator generation frameworks such as FINN and hls4ml. In this article, we present an alternative backend library for FINN, leveraging RTL. We investigate and evaluate, across a spectrum of design dimensions, the pros and cons of an RTL-based implementation versus the original HLS variant. We show that for smaller design parameters, RTL produces significantly smaller circuits as compared with HLS. For larger circuits, however, the look-up table (LUT) count of RTL-based design is slightly higher, up to around 15%. On the other hand, HLS consistently requires more flip-flops (FFs; with an orders-of-magnitude difference for smaller designs) and block RAMs (BRAMs; 2× more). This also impacts the critical path delay, with RTL producing significantly faster circuits, up to around 80%. RTL also benefits from at least a 10× reduction in synthesis time. Finally, the results were validated in practice using two real-world use cases, one of a multi-layer perceptron (MLP) used in network intrusion detection and the other a convolution network called ResNet, used in image recognition. Overall, since HLS frameworks code-generate the hardware design, the benefits of the ease in the design entry is less important. As such, the gained benefits in synthesis time together with some design-dependent resource benefits make the RTL abstraction an attractive alternative.

Customized compute acceleration in the datacenter is key to the wider roll-out of applications based on deep neural network (DNN) inference. In this article, we investigate how to maximize the performance and scalability of field-programmable gate array (FPGA)-based pipeline dataflow DNN inference accelerators (DFAs) automatically on computing infrastructures consisting of multi-die, network-connected FPGAs. We present Elastic-DF, a novel resource partitioning tool and associated FPGA runtime infrastructure that integrates with the DNN compiler FINN. Elastic-DF allocates FPGA resources to DNN layers and layers to individual FPGA dies to maximize the total performance of the multi-FPGA system. In the resulting Elastic-DF mapping, the accelerator may be instantiated multiple times, and each instance may be segmented across multiple FPGAs transparently, whereby the segments communicate peer-to-peer through 100 Gbps Ethernet FPGA infrastructure, without host involvement. When applied to ResNet-50, Elastic-DF provides a 44% latency decrease on Alveo U280. For MobileNetV1 on Alveo U200 and U280, Elastic-DF enables a 78% throughput increase, eliminating the performance difference between these cards and the larger Alveo U250. Elastic-DF also increases operating frequency in all our experiments, on average by over 20%. Elastic-DF therefore increases performance portability between different sizes of FPGA and increases the critical throughput per cost metric of datacenter inference.

Numerous algorithmic optimization techniques have been proposed to alleviate the computational complexity of convolutional neural networks. Given the broad selection of AI accelerators, it is not obvious which approach benefits from which optimization most. The design space includes a large number of deployment settings (batch sizes, power modes, etc.) and unclear measurement methods. This research provides clarity into this design space, leveraging a novel benchmarking approach. We provide a theoretical evaluation of different CNNs and hardware platforms, focusing on understanding the impact of pruning and quantization as primary optimization techniques. We benchmark across a spectrum of FPGA, GPU, TPU, and VLIW processors for systematically pruned and quantized neural networks (ResNet50, GoogLeNetv1, MobileNetv1, a VGG derivative, a multilayer perceptron) over many deployment options, considering power, latency, and throughput at a specific accuracy. Our findings show that channel pruning is most effective and works across most hardware platforms, with speedups directly correlated to the reduction in compute load, while FPGAs benefit the most from quantization. Pruning and quantization are orthogonal, and yield optimal design points when combined. Further in-depth results can be found at our web portal, where we share all experimental data, provide data analytics, and invite the community to contribute.

Deep Neural Networks (DNNs) have a wide application scope beyond computer vision tasks, promising to replace manual algorithmic implementations in applications ranging from large-scale physics experiments to next-generation network security. Such applications may require data processing rates in the millions of samples per second and sub-microsecond latency, which is possible with customized FPGA or ASIC implementations. We present a novel method called LogicNets for co-design of DNN topologies and hardware circuits that maps to a very efficient FPGA implementation to address the needs of such applications.

Numerous algorithmic optimization techniques have been proposed to alleviate the computational complexity of convolutional neural networks (CNNs). However, given the broad selection of inference accelerators, it is not obvious which approach benefits from which optimization and to what degree. In addition, the design space is further obscured by many deployment settings such as power and operating modes, batch sizes, as well as ill-defined measurement methodologies. In this paper, we systematically benchmark different types of CNNs leveraging both pruning and quantization as the most promising optimization techniques leveraging a novel benchmarking approach. We evaluate a spectrum of FPGA implementations, GPU, TPU and VLIW processor, for a selection of systematically pruned and quantized neural networks (including ResNet50, GoogleNetv1, MobileNetv1, a VGG derivative, and a multilayer perceptron) taking the full design space into account including batch sizes, thread counts, stream sizes and operating modes, and considering power, latency, and throughput at a specific accuracy as figure of merit. Our findings show that channel pruning is effective across most hardware platforms, with resulting speedups directly correlated to the reduction in compute load, while FPGAs benefit the most from quantization. FPGAs outperform regarding latency and latency variation for the majority of CNNs, in particular with feed-forward dataflow implementations. Finally, pruning and quantization are orthogonal techniques and yield the majority of all optimal design points when combined. With this benchmarking approach, both in terms of methodology and measured results, we aim to drive more clarity in the choice of CNN implementations and optimizations.

Low bit quantization of neural network is required on edge devices to achieve lower power consumption and higher performance. 8bit or binary network either consumes a lot of resources or has accuracy degradation. Thus, a full-process hardware-friendly quantization solution of 4A4W (activations 4bit and weights 4bit) is proposed to achieve better accuracy/resource trade-off. It doesn't contain any additional floating operations and achieve accuracy comparable to full-precision. We also implement a low-precision accelerator for CNN (LPAC) on the Xilinx FPGA, which takes full advantage of its DSP by efficiently mapping convolutional computations. Through on-chip reassign management and resource-saving analysis, high performance can be achieved on small chips. Our 4A4W solution achieves 1.8x higher performance than 8A8W and 2.42x increase in power efficiency under the same resource. On ImageNet classification, the accuracy has a gap less than 1% to full-precision in Top-5. On the human pose estimation, we achieve 261 frames per second on ZU2EG, which is 1.78x speed up compared to 8A8W and the accuracy has only 1.62% gap to full-precision. This proves that our solution has better universality.

Deployment of deep neural networks for applications that require very high throughput or extremely low latency is a severe computational challenge, further exacerbated by inefficiencies in mapping the computation to hardware. We present a novel method for designing neural network topologies that directly map to a highly efficient FPGA implementation. By exploiting the equivalence of artificial neurons with quantized inputs/outputs and truth tables, we can train quantized neural networks that can be directly converted to a netlist of truth tables, and subsequently deployed as a highly pipelinable, massively parallel FPGA circuit. However, the neural network topology requires careful consideration since the hardware cost of truth tables grows exponentially with neuron fan-in. To obtain smaller networks where the whole netlist can be placed-and-routed onto a single FPGA, we derive a fan-in driven hardware cost model to guide topology design, and combine high sparsity with few-bit activation quantization to limit the neuron fan-in. We evaluate our approach on two tasks with very high intrinsic throughput requirements in high-energy physics and network intrusion detection. We show that the combination of sparsity and few-bit activation quantization results in high-speed circuits with small logic depth and low LUT cost, demonstrating competitive accuracy with less than 15 ns of inference latency and throughput in the hundreds of millions of inferences per second.

Convolutional Neural Network (CNN) dataflow inference accelerators implemented in Field-Programmable Gate Arrays (FPGAs) have demonstrated increased energy efficiency and lower latency compared to CNN execution on CPUs or GPUs. However, the complex shapes of CNN parameter memories do not typically map well to FPGA On-Chip Memories (OCM), which results in poor OCM utilization and ultimately limits the size and types of CNNs which can be effectively accelerated on FPGAs. In this work, we present a design methodology that improves the mapping efficiency of CNN parameters to FPGA OCM. We frame the mapping as a bin packing problem and determine that traditional bin packing algorithms are not well suited to solve the problem within FPGA- and CNN-specific constraints. We hybridize genetic algorithms and simulated annealing with traditional bin packing heuristics to create flexible mappers capable of grouping parameter memories such that each group optimally fits FPGA on-chip memories. We evaluate these algorithms on a variety of FPGA inference accelerators. Our hybrid mappers converge to optimal solutions in a matter of seconds for all CNN use-cases, achieve an increase of up to 65% in OCM utilization efficiency for deep CNNs, and are up to 200× faster than current state-of-the-art simulated annealing approaches.

Deep neural networks (DNNs) are state-of-the-art algorithms for multiple applications, spanning from image classification to speech recognition. While providing excellent accuracy, they often have enormous compute and memory requirements. As a result of this, quantized neural networks (QNNs) are increasingly being adopted and deployed especially on embedded devices, thanks to their high accuracy, but also since they have significantly lower compute and memory requirements compared to their floating point equivalents. QNN deployment is also being evaluated for safety-critical applications, such as automotive, avionics, medical or industrial. These systems require functional safety, guaranteeing failure-free behaviour even in the presence of hardware faults. In general fault tolerance can be achieved by adding redundancy to the system, which further exacerbates the overall computational demands and makes it difficult to meet the power and performance requirements. In order to decrease the hardware cost for achieving functional safety, it is vital to explore domain-specific solutions which can exploit the inherent features of DNNs. In this work we present a novel methodology called fault-aware training (FAT), which includes error modeling during neural network (NN) training, to make QNNs resilient to specific fault models on the device. Our experiments show that by injecting faults in the convolutional layers during training, highly accurate convolutional neural networks (CNNs) can be trained which exhibits much better error tolerance compared to the original. Furthermore, we show that redundant systems which are built from QNNs trained with FAT achieve higher worse-case accuracy at lower hardware cost. This has been validated for numerous classification tasks including CIFAR10, GTSRB, SVHN and ImageNet.

Using FPGAs to accelerate ConvNets has attracted significant attention in recent years. However, FPGA accelerator design has not leveraged the latest progress of ConvNets. As a result, the key application characteristics such as frames-per-second (FPS) are ignored in favor of simply counting GOPs, and results on accuracy, which is critical to application success, are often not even reported. In this work, we adopt an algorithm-hardware co-design approach to develop a ConvNet accelerator called Synetgy and a novel ConvNet model called DiracDeltaNet. Both the accelerator and ConvNet are tailored to FPGA requirements. DiracDeltaNet, as the name suggests, is a ConvNet with only $1\times 1$ convolutions while spatial convolutions are replaced by more efficient shift operations. DiracDeltaNet achieves competitive accuracy on ImageNet (89.0% top-5), but with 48× fewer parameters and 65× fewer OPs than VGG16. We further quantize DiracDeltaNet's weights to 1-bit and activations to 4-bits, with less than 1% accuracy loss. These quantizations exploit well the nature of FPGA hardware. In short, DiracDeltaNet's small model size, low computational OP count, ultra-low precision and simplified operators allow us to co-design a highly customized computing unit for an FPGA. We implement the computing units for DiracDeltaNet on an Ultra96 SoC system through high-level synthesis. Our accelerator's final top-5 accuracy of 88.2% on ImageNet, is higher than all the previously reported embedded FPGA accelerators. In addition, the accelerator reaches an inference speed of 96.5 FPS on the ImageNet classification task, surpassing prior works with similar accuracy by at least 16.9×.

Neural Networks have become one of the most successful universal machine-learning algorithms. They play a key role in enabling machine vision and speech recognition and are increasingly adopted in other application domains. Their computational complexity is enormous and comes along with equally challenging memory requirements in regards to capacity and access bandwidth, which limits deployment in particular within energy constrained, embedded environments. To address these implementation challenges, a broad spectrum of new customized and heterogeneous hardware architectures have emerged, often accompanied with co-designed algorithms to extract maximum benefit out of the hardware. Furthermore, numerous optimization techniques are being explored for neural networks to reduce compute and memory requirements while maintaining accuracy. This results in an abundance of algorithmic and architectural choices, some of which fit specific use cases better than others. For system-level designers, there is currently no good way to compare the variety of hardware, algorithm, and optimization options. While there are many benchmarking efforts in this field, they cover only subsections of the embedded design space. None of the existing benchmarks support essential algorithmic optimizations such as quantization, an important technique to stay on chip, or specialized heterogeneous hardware architectures. We propose a novel benchmark suite, QuTiBench , that addresses this need. QuTiBench is a novel multi-tiered benchmarking methodology ( Ti ) that supports algorithmic optimizations such as quantization ( Qu ) and helps system developers understand the benefits and limitations of these novel compute architectures in regard to specific neural networks and will help drive future innovation. We invite the community to contribute to QuTiBench to support the full spectrum of choices in implementing machine-learning systems.

Inference for state-of-the-art deep neural networks is computationally expensive, making them difficult to deploy on constrained hardware environments. An efficient way to reduce this complexity is to quantize the weight parameters and/or activations during training by approximating their distributions with a limited entry codebook. For very low-precisions, such as binary or ternary networks with 1-8-bit activations, the information loss from quantization leads to significant accuracy degradation due to large gradient mismatches between the forward and backward functions. In this paper, we introduce a quantization method to reduce this loss by learning a symmetric codebook for particular weight subgroups. These subgroups are determined based on their locality in the weight matrix, such that the hardware simplicity of the low-precision representations is preserved. Empirically, we show that symmetric quantization can substantially improve accuracy for networks with extremely low-precision weights and activations. We also demonstrate that this representation imposes minimal or no hardware implications to more coarse-grained approaches. Source code is available at https://www.github.com/julianfaraone/SYQ.

Neural networks have established as a generic and powerful means to approach challenging problems such as image classification, object detection or decision making. Their successful employment foots on an enormous demand of compute. The quantization of network parameters and the processed data has proven a valuable measure to reduce the challenges of network inference so effectively that the feasible scope of applications is expanded even into the embedded domain. This paper describes the making of a real-time object detection in a live video stream processed on an embedded all-programmable device. The presented case illustrates how the required processing is tamed and parallelized across both the CPU cores and the programmable logic and how the most suitable resources and powerful extensions, such as NEON vectorization, are leveraged for the individual processing steps. The crafted result is an extended Darknet framework implementing a fully integrated, end-to-end solution from video capture over object annotation to video output applying neural network inference at different quantization levels running at 16 frames per second on an embedded Zynq UltraScale+ (XCZU3EG) platform.

In this paper, we argue that instead of solely focusing on developing efficient architectures to accelerate well-known low-precision CNNs, we should also seek to modify the network to suit the FPGA. We develop a fully automative toolflow which focuses on modifying the network through filter pruning, such that it efficiently utilizes the FPGA hardware whilst satisfying a predefined accuracy threshold. Although fewer weights are re-moved in comparison to traditional pruning techniques designed for software implementations, the overall model complexity and feature map storage is greatly reduced. We implement the AlexNet and TinyYolo networks on the large-scale ImageNet and PascalVOC datasets, to demonstrate up to roughly 2× speedup in frames per second and 2× reduction in resource requirements over the original network, with equal or improved accuracy.

It is well known that many types of artificial neural networks, including recurrent networks, can achieve a high classification accuracy even with low-precision weights and activations. The reduction in precision generally yields much more efficient hardware implementations in regards to hardware cost, memory requirements, energy, and achievable throughput. In this paper, we present the first systematic exploration of this design space as a function of precision for Bidirectional Long Short-Term Memory (BiLSTM) neural network. Specifically, we include an in-depth investigation of precision vs. accuracy using a fully hardware-aware training flow, where during training quantization of all aspects of the network including weights, input, output and in-memory cell activations are taken into consideration. In addition, hardware resource cost, power consumption and throughput scalability are explored as a function of precision for FPGA-based implementations of BiLSTM, and multiple approaches of parallelizing the hardware. We provide the first open source HLS library extension of FINN for parameterizable hardware architectures of LSTM layers on FPGAs which offers full precision flexibility and allows for parameterizable performance scaling offering different levels of parallelism within the architecture. Based on this library, we present an FPGA-based accelerator for BiLSTM neural network designed for optical character recognition, along with numerous other experimental proof points for a Zynq UltraScale+ XCZU7EV MPSoC within the given design space.

Massive spatial parallelism at low energy gives FPGAs the potential to be core components in large scale high performance computing (HPC) systems. In this paper we present four major design steps that harness high-level synthesis (HLS) to implement scalable spatial FPGA algorithms. To aid productivity, we introduce the open source library hlslib to complement HLS. We evaluate kernels designed with our approach on an FPGA accelerator board, demonstrating high performance and board utilization with enhanced programmer productivity. By following our guidelines, programmers can use HLS to develop efficient parallel algorithms for FPGA, scaling their implementations with increased resources on future hardware.

Convolutional Neural Networks have rapidly become the most successful machine-learning algorithm, enabling ubiquitous machine vision and intelligent decisions on even embedded computing systems. While the underlying arithmetic is structurally simple, compute and memory requirements are challenging. One of the promising opportunities is leveraging reduced-precision representations for inputs, activations, and model parameters. The resulting scalability in performance, power efficiency, and storage footprint provides interesting design compromises in exchange for a small reduction in accuracy. FPGAs are ideal for exploiting low-precision inference engines leveraging custom precisions to achieve the required numerical accuracy for a given application. In this article, we describe the second generation of the FINN framework, an end-to-end tool that enables design-space exploration and automates the creation of fully customized inference engines on FPGAs. Given a neural network description, the tool optimizes for given platforms, design targets, and a specific precision. We introduce formalizations of resource cost functions and performance predictions and elaborate on the optimization algorithms. Finally, we evaluate a selection of reduced precision neural networks ranging from CIFAR-10 classifiers to YOLO-based object detection on a range of platforms including PYNQ and AWS F1, demonstrating new unprecedented measured throughput at 50 TOp/s on AWS F1 and 5 TOp/s on embedded devices.

The goal of reaching exascale computing is made especially challenging by the highly heterogeneous nature of modern platforms and the energy they consume. As compute nodes typically utilize multiple multi-core CPU and are increasingly equipped with PCIe based accelerators, both are contributing to an ever more dynamic power consumption. In our study we evaluate our target application on a variety of heterogeneous platforms, including high end FPGA, GPU, and Xeon Phi accelerators, with respect to energy efficiency at a node and cluster level. We compare multiple implementations of our application, each built with a different modern parallel programming framework, with respect to execution performance, code complexity and energy efficiency. Later we extrapolate based on our findings, the implications of scaling this application towards exascale, with projections of computation achievable within the exascale power budget for our three architectures.

Smith-Waterman is a dynamic programming algorithm that plays a key role in the modern genomics pipeline as it is guaranteed to find the optimal local alignment between two strings of data. The state of the art presents many hardware acceleration solutions that have been implemented in order to exploit the high degree of parallelism available in this algorithm. The majority of these implementations use heuristics to increase the performance of the system at the expense of the accuracy of the result. In this work, we present an implementation of the pure version of the algorithm. We include the key architectural optimizations to achieve highest possible performance for a given platform and leverage the Berkeley roofline model to track the performance and guide the optimizations. To achieve scalability, our custom design comprises of systolic arrays, data compression features and shift registers, while a custom port mapping strategy aims to maximize performance. Our designs are built leveraging an OpenCL-based design entry, namely Xilinx SDAccel, in conjunction with a Xilinx Virtex 7 and Kintex Ultrascale platform. Our final design achieves a performance of 42.47 GCUPS (giga cell updates per second) with an energy efficiency of 1.6988 GCUPS/W. This represents an improvement of 1.72x in performance and energy efficiency over previously published FPGA implementations and 8.49x better in energy efficiency over comparable GPU implementations.

Convolutional Neural Networks have dramatically improved in recent years, surpassing human accuracy on certain problems and performance exceeding that of traditional computer vision algorithms. While the compute pattern in itself is relatively simple, significant compute and memory challenges remain as CNNs may contain millions of floating-point parameters and require billions of floating-point operations to process a single image. These computational requirements, combined with storage footprints that exceed typical cache sizes, pose a significant performance and power challenge for modern compute architectures. One of the promising opportunities to scale performance and power efficiency is leveraging reduced precision representations for all activations and weights as this allows to scale compute capabilities, reduce weight and feature map buffering requirements as well as energy consumption. While a small reduction in accuracy is encountered, these Quantized Neural Networks have been shown to achieve state-of-the-art accuracy on standard benchmark datasets, such as MNIST, CIFAR-10, SVHN and even ImageNet, and thus provide highly attractive design trade-offs. Current research has focused mainly on the implementation of extreme variants with full binarization of weights and or activations, as well typically smaller input images. Within this paper, we investigate the scalability of dataflow architectures with respect to supporting various precisions for both weights and activations, larger image dimensions, and increasing numbers of feature map channels. Key contributions are a formalized approach to understanding the scalability of the existing hardware architecture with cost models and a performance prediction as a function of the target device size. We provide validating experimental results for an ImageNet classification on a server class platform, namely the AWS F1 node.

TCP/IP is the predominant communication protocol in modern networks but also one of the most demanding. Consequently, TCP/IP offload is becoming increasingly popular with standard network interface cards. TCP/IP Offload Engines have also emerged for FPGAs, and are being offered by vendors such as Intilop, Fraunhofer HHI, PLDA and Dini Group. With the target application being high-frequency trading, these implementations focus on low latency and support a limited session count. However, many more applications beyond high-frequency trading can potentially be accelerated inside an FPGA once TCP with high session count is available inside the fabric. This way, a network-attached FPGA on ingress and egress to a CPU can accelerate functions such as encryption, compression, memcached and many others in addition to running the complete network stack. This paper introduces a novel architecture for a 10Gbps line-rate TCP/IP stack for FPGAs that can scale with the number of sessions and thereby addresses these new applications. We prototyped the design on a VC709 development board, demonstrating compatibility with existing network infrastructure, operating at full 10Gbps throughput full-duplex while supporting 10,000 sessions. Finally, the design has been described primarily using high-level synthesis, which accelerates development time and improves maintainability.

FPGA-based data processing is becoming increasingly relevant in data centers, as the transformation of existing applications into dataflow architectures can bring significant throughput and power benefits. Furthermore, a tighter integration of computing and network is appealing, as it overcomes traditional bottlenecks between CPUs and network interfaces, and dramatically reduces latency. In this article, we present the design of a novel hash table, a fundamental building block used in many applications, to enable data processing on FPGAs close to the network. We present a fully pipelined design capable of sustaining consistent 10Gbps line-rate processing by deploying a concurrent mechanism to handle hash collisions. We address additional design challenges such as support for a broad range of key sizes without stalling the pipeline through careful matching of lookup time with packet reception time. Finally, the design is based on a scalable architecture that can be easily parameterized to work with different memory types operating at different access speeds and latencies. We have tested the proposed hash table in an FPGA-based memcached appliance implementing a main-memory key-value store in hardware. The hash table is used to index 2 million entries in 24GB of external DDR3 DRAM while sustaining 13 million requests per second, the maximum packet rate that can be achieved with UDP packets on a 10Gbps link for this application.

To make networks more reliable, enormous resources are poured into all phases of the network-equipment lifecycle. The process starts early in the design phase when simulation is used to verify the correctness of a design, and continues through manufacturing and perhaps months of rigorously trials. With over 7,000 Internet RFCs and hundreds of IEEE standards, a typical piece of networking equipment undergoes hundreds of conformance tests before being deployed. Finally, when deployed in a production network, the equipment is tested regularly. Throughout the process, a relentless battery of tests and measurement help ensure the correct operation of the equipment.

Common web infrastructure relies on distributed main memory key-value stores to reduce access load on databases, thereby improving both performance and scalability of web sites. As standard cloud servers provide sub-linear scalability and reduced power efficiency to these kinds of scale-out workloads, we have investigated a novel dataflow architecture for key-value stores with the aid of FPGAs which can deliver consistent 10Gbps throughput. In this paper, we present the design of a novel hash table which forms the centre piece of this dataflow architecture. The fully pipelined design can sustain consistent 10Gbps line-rate performance by deploying a concurrent mechanism to handle hash collisions. We address problems such as support for a broad range of key sizes without stalling the pipeline through careful matching of lookup time with packet reception time. Finally, the design is based on a scalable architecture that can be easily parametrized to work with different memory types operating at different access speeds and latencies. We deployed this hash table in a memcached prototype to index 2 million entries in 24GBytes of external DDR3 DRAM while sustaining 13 million requests per second for UDP binary encoded memcached packets which is the maximum packet rate that can be achieved with memcached on a 10Gbps link.

The fine-grained parallelism inherent in FPGAs has encouraged their use in packet processing systems. To facilitate debugging and performance evaluation, designers require on-chip monitors that provide abstractions of low-level details and a system-level perspective. In this paper, we present five architectures that permit transaction-based communication-centric monitoring of packet processing systems. We compare the resource requirements and filtering functionality of each architecture, demonstrating that sequential matching is more resource efficient than parallel matching. We also show that generic filtering has a low overhead compared to specialised filtering while providing additional flexibility. A scalable architecture is also presented, which is more flexible and adaptable to matching requirements than other architectures. These monitoring architectures permit the implementation of a highly effective test system which provides a system-level perspective and is more resource efficient than conventional RTL debug environments.

The fine-grained parallelism inherent in FPGAs has encouraged their use in packet processing systems. Debugging and performance evaluation of such complex designs can be significantly improved through debug information that provides a system-level perspective and hides the complexity of signal-level debugging. In this paper we present a debugging system that permits transaction-based communication-centric monitoring of packet processing systems. We demonstrate, using two different examples, how this system can improve the debugging information and abstract lower level detail. Furthermore, we demonstrate that transaction monitoring systems require fewer resources than conventional RTL debugging systems and can provide a system-level perspective not permitted by traditional tools.