Object detection and classification is a key task in many computer vision applications such as smart surveillance and autonomous vehicles. Recent advances in deep learning have significantly improved the quality of results achieved by these systems, making them more accurate and reliable in complex environments. Modern object detection systems make use of lightweight convolutional neural networks (CNNs) for feature extraction, coupled with single-shot multi-box detectors (SSDs) that generate bounding boxes around the identified objects along with their classification confidence scores. Subsequently, a non-maximum suppression (NMS) module removes any redundant detection boxes from the final output. Typical NMS algorithms must wait for all box predictions to be generated by the SSD-based feature extractor before processing them. This sequential dependency between box predictions and NMS results in a significant latency overhead and degrades the overall system throughput, even if a high-performance CNN accelerator is used for the SSD feature extraction component. In this paper, we present a novel pipelined NMS algorithm that eliminates this sequential dependency and associated NMS latency overhead. We then use our novel NMS algorithm to implement an end-to-end fully pipelined FPGA system for low-latency SSD-MobileNet-V1 object detection. Our system, implemented on an Intel Stratix 10 FPGA, runs at 400 MHz and achieves a throughput of 2,167 frames per second with an end-to-end batch-1 latency of 2.13 ms. Our system achieves 5.3× higher throughput and 5× lower latency compared to the best prior FPGA-based solution with comparable accuracy.

Deep neural networks (DNNs) have become key solutions in the natural language processing (NLP) domain. However, the existing accelerators customized for their narrow target models cannot support diverse NLP models. Therefore, naively running complex NLP models on the existing accelerators often leads to very marginal performance improvements. For these reasons, architects are now in dire need of a new accelerator that can run various NLP models while taking its full performance potential. In this article, we propose FlexRun, an FPGA-based modular accelerator to efficiently support diverse and complex NLP models. First, we identify key components commonly used by NLP models and implement them on top of a current state-of-the-art FPGA-based accelerator. Next, FlexRun conducts an in-depth design space exploration to find the best accelerator architecture for a target NLP model. Last, FlexRun automatically reconfigures the accelerator based on the exploration results. Our FlexRun design outperforms the current state-of-the-art FPGA-based accelerator by 1.21×–2.73× and 1.15×–1.50× for BERT and GPT2, respectively. Compared to Nvidia’s V100 GPU, FlexRun achieves 2.69× higher performance on average for various BERT and GPT2 models.

FPGAs have traditionally been successful in accelerating stream processing applications where the amount and type of work performed on each record—e.g., image, packet—do not depend on the record's contents. On the other hand, accelerating ‘input-dependent’ stream processing on FPGAs presents a much more challenging problem where different records in the stream can require widely different operations. It is inefficient and unnecessary to support all operations at the same throughput when the distributions of the operations are skewed. In this paper, we examine the application of the ”make the common case fast” strategy to efficiently accelerate ‘input-dependent’ stream processing on FPGAs. In particular, we study the use offast-slow pathandearly-exittechniques in the design of an FPGA-accelerated network intrusion prevention system (IPS). To avoid overfitting when common-case behavior is varied, we further examinecompile-time re-tuningandruntime adaptationtechniques. A quantitative analysis shows that fast-slow path and early-exit techniques can save an order of magnitude of resources compared to a common-case unaware IPS design. Compile-time re-tuning to specific conditions further achieves 30% – 94% BRAM savings relative to a generalized design. Adding runtime adaptation improves the zero-loss throughput by 1.43 – 2.75 × compared to a fixed design.

This work introduces Remarn, a reconfigurable multi-threaded multi-core accelerator supporting both spatial and temporal co-execution of Recurrent Neural Network (RNN) inferences. It increases processing capabilities and quality of service of cloud-based neural processing units (NPUs) by improving their hardware utilization and by reducing design latency, with two innovations. First, a custom coarse-grained multi-threaded RNN/Long Short-Term Memory (LSTM) hardware architecture, switching tasks among threads when RNN computational engines meet data hazards. Second, the partitioning of this hardware architecture into multiple full-fledged sub-accelerator cores, enabling spatially co-execution of multiple RNN/LSTM inferences. These innovations improve the exploitation of the available parallelism to increase runtime hardware utilization and boost design throughput. Evaluation results show that a dual-threaded quad-core Remarn NPU achieves 2.91 times higher performance while only occupying 5.0% more area than a single-threaded one on a Stratix 10 FPGA. When compared with a Tesla V100 GPU implementation, our design achieves 6.5 times better performance and 15.6 times higher power efficiency, showing that our approach contributes to high performance and energy-efficient FPGA-based multi-RNN inference designs for datacenters.

Current FPGA programming tools require extensive hardware-specific manual code tuning to achieve performance, which is intractable for most software application teams. We present HPVM2FPGA, a novel end-to-end compiler and auto-tuning system that can automatically tune hardware-agnostic programs for FPGAs. HPVM2FPGA uses a hardware-agnostic abstraction of parallelism as an intermediate representation (IR) to represent hardware-agnostic programs. HPVM2FPGA's powerful optimization framework uses sophisticated compiler optimizations and design space exploration (DSE) to automatically tune a hardware-agnostic program for a given FPGA. HPVM2FPGA is able to support software programmers by shifting the burden of performing hardware-specific optimizations to the compiler and DSE. We show that HPVM2FPGA can achieve up to 33×speedup compared to unoptimized baselines and can match the performance of hand-tuned HLS code for three of four benchmarks. We have designed HPVM2FPGA to be a modular and extensible framework, and we expect it to match hand-tuned code for most programs as the system matures with more optimizations. Overall, we believe that it constitutes a solid step closer to fully hardware-agnostic FPGA programming, making it a suitable cornerstone for future FPGA compiler research.

With the continued growth in field-programmable gate array (FPGA) capacity and their incorporation into new environments such as datacenters, we have witnessed the introduction of a new class of reconfigurable acceleration devices (RADs) that go beyond conventional FPGA architectures. These devices combine a reconfigurable fabric with coarse-grained domain-specialized accelerator blocks all connected via a high-performance packet-switched network-on-chip (NoC) for efficient system-wide communication. However, we lack the tools necessary to efficiently explore the huge design space for RADs, study the complex interactions between their different components and evaluate various combinations of design choices. In this work, we develop RAD-Sim, a cycle-level architecture simulator that allows rapid application-driven exploration of the design space of novel RADs. To showcase the capabilities of RAD-Sim, we map and simulate a state-of-the-art deep learning (DL) inference overlay on a RAD instance incorporating an FPGA fabric and a complex of hard matrix-vector multiplication engines, communicating over a system-wide NoC. Through this example, we show how RAD-Sim can help architects quantify the effect of changing specific architecture parameters on end-to-end application performance.

The advent of AI has driven the exploration of high-density low-precision arithmetic on FPGAs. This has resulted in new methods in mapping both arithmetic functions as well as dataflows onto the fabric, as well as some changes to the embedded DSP Blocks. Technologies outside of the FPGA realm have also evolved, such as the addition of tensor structures for GPUs, as well as the introduction of numerous AI ASSPs, all of which have a higher claimed performance and efficiency than current FPGAs. In this article, we will introduce the Stratix 10 NX device, which is a variant of FPGA specifically optimized for the AI application space. In addition to the computational capabilities of the standard programmable soft-logic fabric, a new type of DSP Block provides the dense arrays of low-precision multipliers typically used in AI implementations. The architecture of the block is tuned for the common matrix-matrix or vector-matrix multiplications in AI, with capabilities designed to work efficiently for both small and large matrix sizes. The base precisions are INT8 and INT4, along with shared exponent support to support block FP16 and block FP12 numerics. All additions/accumulations can be done in INT32 or IEEE-754 single precision floating point (FP32), and multiple blocks can be cascaded together to support larger matrices. We will also describe methods by which the smaller precision multipliers can be aggregated to create larger multipliers that are more applicable to standard signal processing requirements. In the AI market, the FPGA must compete directly with other types of devices, rather than occupy a unique niche. Deterministic system performance is as important as the performance of individual FPGA elements, such as logic, memory, and DSP. We will show that the feed forward datapath structures that are needed to support the typical AI matrix-vector and matrix-matrix multiplication operations can consistently close timing at over 500 MHz on a mid-speed grade device, even if all of the Tensor Blocks on the device are used. We will also show a full-chip NPU processor implementation that out performs GPUs at the same process node for a variety of AI inferencing workloads, even though it has a lower operating frequency of 365 MHz. In terms of overall compute throughput, Stratix 10 NX is specified at 143 INT8/FP16 TOPs/FLOPs or 286 INT4/FP12 TOPS/FLOPs. Depending on the configuration, power efficiency is in the range of 1–4 TOPs or TFLOPs/W.

This article presents a reconfigurable accelerator for REcurrent Neural networks with fine-grained cOlumn-Wise matrix–vector multiplicatioN (RENOWN). We propose a novel latency-hiding architecture for recurrent neural network (RNN) acceleration using column-wise matrix–vector multiplication (MVM) instead of the state-of-the-art row-wise operation. This hardware (HW) architecture can eliminate data dependencies to improve the throughput of RNN inference systems. Besides, we introduce a configurable checkerboard tiling strategy which allows large weight matrices, while incorporating various configurations of element-based parallelism (EP) and vector-based parallelism (VP). These optimizations improve the exploitation of parallelism to increase HW utilization and enhance system throughput. Evaluation results show that our design can achieve over 29.6 tera operations per second (TOPS) which would be among the highest for field-programmable gate array (FPGA)-based RNN designs. Compared to state-of-the-art accelerators on FPGAs, our design achieves 3.7–14.8 times better performance and has the highest HW utilization.

The density of FPGA on-chip memory has been continuously increasing with modern FPGAs having thousands of block RAMs (BRAMs) distributed across their reconfigurable fabric. These distributed BRAMs can provide a tremendous amount of on-chip bandwidth for efficient acceleration of data-intensive applications. In this work, we propose enhancing the ubiquitous FPGA BRAMs with in-memory compute-capabilities. As a result, BRAMs can act as normal storage units or their bitlines can be re-purposed as SIMD lanes executing bit-serial arithmetic operations. Our proposed architectural change results in 1.6× and 2.3× increase in the peak multiply-accumulate throughput of a large Stratix 10 FPGA, at a minimal cost of only 1.8% increase in the FPGA die size and no change to the BRAM's interface to the programmable routing. Then, we present RIMA, a reconfigurable in-memory accelerator architecture for deep learning (DL) inference. RIMA exploits the proposed compute-capable BRAMs and the FPGA's reconfigurability to achieve 1.25× and 3× higher performance compared to the state-of-the-art Brainwave DL soft processor for 8-bit integer and block floating-point precisions, respectively. In addition, RIMA implemented on a Stratix 10 FPGA enhanced with compute-capable BRAMs can achieve an order of magnitude higher performance compared to a same-generation GPU.

The advent of AI has driven the adoption of high density low precision arithmetic on FPGAs. This has resulted in new methods in mapping both arithmetic functions as well as dataflows onto the fabric, as well as some changes to the embedded DSP Blocks. Technologies outside of the FPGA realm have also evolved, such as the addition of tensor structures for GPUs, and also the introduction of numerous AI ASSPs, all of which have a higher claimed performance and efficiency than current FPGAs. In this paper we will introduce the Stratix 10 NX device (NX), which is a variant of FPGA specifically optimized for the AI application space. In addition to the computational capabilities of the standard programmable soft logic fabric, a new type of DSP Block provides the dense arrays of low precision multipliers typically used in AI implementations. The architecture of the block is tuned for the common matrix-matrix or vector-matrix multiplications in AI, with capabilities designed to work efficiently for both small and large matrix sizes. The base precisions are INT8 and INT4, along with shared exponent support for support block floating point FP16 and FP12 numerics. All additions/accumulations can be done in INT32 or IEEE754 single precision floating point (FP32), and multiple blocks can be cascaded together to support larger matrices. We will also describe methods by which the smaller precision multipliers can be aggregated to create larger multiplier that are more applicable to standard signal processing requirements. In terms of overall compute throughput, Stratix 10 NX achieves 143 INT8/FP16 TOPs/FLOPs, or 286 INT4/FP12 TOPS/FLOPs at 600MHz. Depending on the configuration, power efficiency is in the range of 1-4 TOPs or TFLOPs/W.

Object detection is an important computer vision task, with many applications in autonomous driving, smart surveillance, robotics, and other domains. Single-shot detectors (SSD) coupled with a convolutional neural network (CNN) for feature extraction can efficiently detect, classify and localize various objects in an input image with very high accuracy. In such systems, the convolution layers extract features and predict the bounding box locations for the detected objects as well as their confidence scores. Then, a non-maximum suppression (NMS) algorithm eliminates partially overlapping boxes and selects the bounding box with the highest score per class. However, these two components are strictly sequential; a conventional NMS algorithm needs to wait for all box predictions to be produced before processing them. This prohibits any overlap between the execution of the convolutional layers and NMS, resulting in significant latency overhead and throughput degradation. In this paper, we present a novel NMS algorithm that alleviates this bottleneck and enables a fully-pipelined hardware implementation. We also implement an end-to-end system for low-latency SSD-MobileNet-V1 object detection, which combines a state-of-the-art deeply-pipelined CNN accelerator with a custom hardware implementation of our novel NMS algorithm. As a result of our new algorithm, the NMS module adds a minimal latency overhead of only 0.13μ s to the SSD-MobileNet-V1 convolution layers. Our end-to-end object detection system implemented on an Intel Stratix 10 FPGA runs at a maximum operating frequency of 350 MHz, with a throughput of 609 frames-per-second and an end-to-end batch-1 latency of 2.4 ms. Our system achieves 1.5× higher throughput and 4.4× lower latency compared to the current state-of-the-art SSD-based object detection systems on FPGAs.

”Soft” GPUs are overlays that implement GPGPU-like data parallel processor architectures in FPGA logic to make FPGAs as software-programmable as ”hard” GPGPUs. Unlike hard GPUs, soft GPU architectures can be specialized to further improve efficiency by leveraging FPGA’s flexibility. Prior work has shown the software programmability potential for soft GPUs but only studied general-purpose soft GPUs with minor specializations (e.g., FPGU, FlexGrip, MIAOW, and SCRATCH) or only domain-optimized for a particular application domain (e.g., PDL-FGPU for the persistent deep learning domain.) This paper proposes a soft GPU development framework to automate the creation of soft GPU instances with aggressive application-domain optimizations (i.e., domain-optimized GPUs, or DOGPUs) that consists of a baseline general soft GPU architecture ”template” with an improved architecture over prior general purpose soft GPUs, along with a customizable partition that enables a custom datapath (macro unit) to be inserted to optimize for a target application domain. Unlike the prior PDL-FGPU which targets the persistent deep learning domain, the proposed framework can be used to target optimization for any application domain. Our evaluation on a set of data parallel workloads shows that (i) the proposed general soft GPU architecture offers average speedup of 1.8x versus the best prior soft GPUs we know of (i.e., FGPU, PDL-FGPU), (ii) DO-GPUs with domain-optimizations provide an average of 218x speedup over general soft GPUs, (iii) the proposed framework enabled building six new domain-optimized soft GPU instances in a matter of days, and (iv) enables quick GPU-like development effort (hours), where code is concise (low 100s of lines) and can be compiled in seconds without FPGA EDA tools in the loop, assuming an appropriate soft DO-GPU bitstream for the application domain is already built.

We present a novel cache compression method that leverages the fine-grained data duplication across cache lines. We leverage the XOR operation of the in-SRAM bit-line computing peripherals, to search for compressible data over a wide range of data locations on cache, reducing the data movement requirements. To reduce the decompression latency, we design specialized compression schemes by fetching the data with the same parallelism as the original cache, according to the architecture of the last-level cache slice. The proposed compression method achieves a 2.05× compression ratio on average (up to 67×), and 4.73% of speedup on average (up to 29%), over the SPEC2006 benchmarks.

In today's increasingly heterogeneous compute landscape, there is high demand for design tools that offer seemingly contradictory features: portable programming abstractions that hide underlying architectural detail, and the capability to optimise and exploit architectural features. Our meta-programming approach, Artisan, decouples application functionality from optimisation concerns to address the complexity of mapping high-level application descriptions onto heterogeneous platforms from which they are abstracted. With Artisan, application experts focus on algorithmic behaviour, while platform and domain experts focus on optimisation and mapping. Artisan offers complete design-flow orchestration in a unified programming environment based on Python 3 to enable accessible codification of reusable optimisation strategies that can be automatically applied to high-level application descriptions. We have developed and evaluated an Artisan prototype and a set of customised meta-programs used to automatically optimise six case study applications for CPU+FPGA targets. In our experiments, Artisan-optimised designs achieve the same order of magnitude speedup as manually optimised designs compared to corresponding unoptimised software.

Overlay architectures are a good way to enable fast development and debug on FPGAs at the expense of potentially limited performance compared to fully customized FPGA designs. When used in concert with hand-tuned FPGA solutions, performant overlay architectures can improve time-to-solution and thus overall productivity of FPGA solutions. This work tunes and specializes FGPU, an open source OpenCL-programmable GPU overlay for FPGAs. We demonstrate that our persistent deep learning (PDL )-FGPU architecture maintains the ease-of-programming and generality of GPU programming while achieving high performance from specialization for the persistent deep learning domain. We also propose an easy method to specialize for other domains. PDL-FGPU includes new instructions, along with micro-architecture and compiler enhancements. We evaluate both the FGPU baseline and the proposed PDL-FGPU on a modern high-end Intel Stratix 10 2800 FPGA in simulation running persistent DL applications (RNN, GRU, LSTM), and non-DL applications to demonstrate generality. PDL-FGPU requires 1.4–3× more ALMs, 4.4–6.4× more M20ks, and 1–9.5× more DSPs than baseline, but improves performance by 56–693× for PDL applications with an average 23.1% degradation on non-PDL applications. We integrated the PDL-FGPU overlay into Intel OPAE to measure real-world performance/power and demonstrate that PDL-FGPU is only 4.0–10.4× slower than the Nvidia V100.

This paper presents SLATE, a fully-managed, heterogeneous Function-as-a-Service (FaaS) system for deploying serverless functions onto heterogeneous cloud infrastructures. We extend the traditional homogeneous FaaS execution model to support heterogeneous functions, automating and abstracting runtime management of heterogeneous compute resources in order to improve cloud tenant accessibility to specialised, accelerator resources, such as FPGAs and GPUs. In particular, we focus on the mechanisms required for heterogeneous scaling of deployed function instances to guarantee latency objectives while minimising cost. We develop a simulator to validate and evaluate our approach, considering case-study functions in three application domains: machine learning, bio-informatics, and physics. We incorporate empirically derived performance models for each function implementation targeting a hardware platform with combined computational capacity of 24 FPGAs and 12 CPU cores. Compared to homogeneous CPU and homogeneous FPGA functions, simulation results achieve respectively a cost improvement for non-uniform task traffic of up to 8.7 times and 1.7 times, while maintaining specified latency objectives.

The increasing size of recurrent neural networks (RNNs) makes it hard to meet the growing demand for real-time AI services. For low-latency RNN serving, FPGA-based accelerators can leverage specialized architectures with optimized dataflow. However, they also suffer from severe HW under-utilization when partitioning RNNs, and thus fail to obtain the scalable performance.In this paper, we identify the performance bottlenecks of existing RNN partitioning strategies. Then, we propose a novel RNN partitioning strategy to achieve the scalable multi-FPGA acceleration for large RNNs. First, we introduce three parallelism levels and exploit them by partitioning weight matrices, matrix/vector operations, and layers. Second, we examine the performance impact of collective communications and software pipelining to derive more accurate and optimal distribution results. We prototyped an FPGA-based acceleration system using multiple Intel high-end FPGAs, and our partitioning scheme allows up to 2.4x faster inference of modern RNN workloads than conventional partitioning methods.

This paper proposes a novel latency-hiding hardware architecture based on column-wise matrix-vector multiplication to eliminate data dependency, improving the throughput of systems of RNN models. In addition, a flexible checkerboard tiling strategy is introduced to allow large weight matrices, while supporting element-based parallelism and vector-based parallelism. These optimizations improve the exploitation of the available parallelism to increase run-time hardware utilization and boost inference throughput. Furthermore, a quantization scheme with fine-tuning is proposed to achieve high accuracy. Evaluation results show that the proposed architecture can enhance performance and energy efficiency with little accuracy loss. It achieves 1.05 to 3.35 times better performance and 1.22 to 3.92 times better hardware utilization than a state-of-theart FPGA-based LSTM design, which shows that our approach contributes to high performance FPGA-based LSTM systems.

This paper provides a novel compilation approach that addresses the complexity of mapping high-level descriptions to heterogeneous platforms, improving design productivity and maintainability. Our approach is based on a co-design methodology decoupling functional concerns from optimisation concerns, allowing two separate descriptions to be independently maintained by two types of programmers: application experts focus on algorithmic behaviour, while platform experts focus on the mapping process. Our approach supports two key requirements: (1) Customisable optimisations to rapidly capture a wide range of mapping strategies, and (2) Reusable strategies to allow optimisations to be described once and applied to multiple applications. To evaluate our approach, we develop Artisan, a meta-programming tool for codifying optimisation strategies using a high-level general-purpose programming language (Python 3), offering full design-flow orchestration of key components (source-code, third-party tools, and platforms). We evaluate Artisan using three case study applications and three reusable optimisation strategies, achieving at least 24 times speedup for each application on CPU and FPGA targets with little application developer effort.

Training convolutional neural networks (CNNs) requires intensive computations as well as a large amount of storage and memory access. While low bandwidth off-chip memories in prior FPGA works have hindered the system-level performance, modern FPGAs offer high bandwidth memory (HBM2) that unlocks opportunities to improve the throughput/energy of FPGA-based CNN training. This paper presents a FPGA accelerator for CNN training which (1) uses HBM2 for efficient off-chip communication, and (2) supports various training operations (e.g. residual connections, stride-2 convolutions) for modern CNNs. We analyze the impact of HBM2 on CNN training workloads, provide a comprehensive comparison with DDR3, and present the strategies to efficiently use HBM2 features for enhanced CNN training performance. For training ResNet-20/VGG-like CNNs for CIFAR-10 dataset with low batch size of 2, the proposed CNN training accelerator on Intel Stratix-10 MX FPGA demonstrates 1.4/1.7X energy-efficiency improvement compared to Stratix-10 GX FPGA with DDR3 memory, and 4.5/9.7 X energy-efficiency improvement compared to Tesla V100 GPU.

Interactive intelligent services, such as smart web search, are important datacenter workloads. They rely on dataintensive deep learning (DL) algorithms with strict latency constraints and thus require balancing both data movement and compute capabilities. As such, a persistent approach that keeps the entire DL model on-chip is becoming the new norm for realtime services to avoid the expensive off-chip memory accesses. This approach is adopted in Microsoft's Brainwave and is also provided by Nvidia's cuDNN libraries. This paper presents a comparative study of FPGA, GPU, and FPGA+ASIC in-package solutions for persistent DL. Unlike prior work, we offer a fair and direct comparison targeting common numerical precisions (FP32, INT8) and modern high-end FPGA (Intel® Stratix®10), GPU (Nvidia Volta), and ASIC (10 nm process), all using the persistent approach. We show that Stratix 10 FPGAs offer 2.7× (FP32) to 8.6× (INT8) lower latency than Volta GPUs across RNN, GRU, and LSTM workloads from DeepBench. The GPU can only utilize ~6% of its peak TOPS, while the FPGA with a more balanced on-chip memory and compute can achieve much higher utilization (~57%). We also study integrating an ASIC chiplet, TensorRAM, with an FPGA as system-in-package to enhance on-chip memory capacity and bandwidth, and provide compute throughput matching the required bandwidth. We show that a small 32 mm2 TensorRAM 10nm chiplet can offer 64 MB memory, 32 TB/s on-chiplet bandwidth, and 64 TOPS (INT8). A small Stratix 10 FPGA with a TensorRAM (INT8) offers 15.9× better latency than GPU (FP32) and 34× higher energy efficiency. It has 2× aggregate on-chip memory capacity compared to a large FPGA or GPU. Overall, our study shows that the FPGA is better than the GPU for persistent DL, and when integrated with an ASIC chiplet, it can offer a more compelling solution.

FPGA-based processing has gained much attention for accelerating graph analytics because of the demand in performance and energy efficiency. However, while priority scheduling has been shown to be an effective optimization for improving performance for worklist-based graph computations, it is rarely used in accelerator designs due to its implementation complexity and memory-access overhead. In this paper, we present a heterogeneous processing approach for priority scheduling on a shared-memory CPU-FPGA platform. By exploiting the closely-coupled integration of the host processor and the FPGA accelerator, our system dynamically offloads the task of scheduling to a software scheduler on the processor for its programmability, high-capacity cache and low memory latency, while the FPGA graph processing accelerator enjoys the scheduling benefit and delivers higher performance at excellent energy efficiency. To understand the effectiveness of our solution, we compared it with FPGA-only solutions for two scheduling schemes: the well-known Dijkstra scheduling for Single Source Shortest Path and a new scheduling optimization we developed for improving the data locality of Breadth First Search. Whereas the FPGA-only solution requires an impractical amount of on-chip storage to implement a priority queue, the proposed processor-assisted scheduling that moves the task of scheduling to the processor consumes a negligible load on the processor and retains most of the performance benefit from priority scheduling.

Interactive intelligent services (e.g., smart web search) are becoming essential datacenter workloads. They rely on data-intensive artificial intelligence (AI) algorithms that do not use batch computation due to their tight latency constraints. Since off-chip data accesses have higher latency and energy consumption than on-chip accesses, a persistent AI approach with the entire model stored in on-chip memory is becoming the new norm for real-time AI. This approach is the cornerstone of Microsoft's Brainwave FPGA-based AI cloud and was recently added to Nvidia's cuDNN library. In this work, we implement, optimize and evaluate a Brainwave-like neural processing unit (NPU) on a large Stratix-10 FPGA. We benchmark it against a large Nvidia Volta GPU running cuDNN persistent AI kernels. Across real-time persistent RNN, GRU, and LSTM workloads, we show that Stratix-10 offers ~3× (FP32) and ~10× (INT8) better latency than GPU (FP32), which uses only ~6% of its peak throughput. Then, we propose TensorRAM, an ASIC chiplet for persistent AI that is 2.5D integrated with an FPGA in the same package. TensorRAM enhances the on-chip memory capacity and bandwidth, with enough multi-precision INT8/4/2/1 throughput to match that bandwidth. Multiple TensorRAMs can be integrated with Stratix-10. Our evaluation shows that a small 32-mm2 TensorRAM on 10nm offers 64MB of SRAMs with 32TB/s on-chiplet bandwidth and 64 TOP/s (INT8). A small Stratix-10 with a TensorRAM (INT8) offers 16× better latency and 34× energy efficiency compared to GPU (FP32). Overall, Stratix-10 with TensorRAM offers compelling and scalable persistent AI solutions.

Overlay architectures are a good way to enable fast development and debug on FPGAs at the expense of potentially limited performance when compared to fully customized FPGA designs. When used in concert with a hand-tuned FPGA solution, a performant overlay architecture can improve the time-to-solution and thus overall productivity of FPGA solutions. In this work, we tune and specialize FGPU, an open source OpenCL-programmable GPU overlay for FPGAs. We demonstrate that our PDL-FGPU architecture is able to maintain the ease-of-programming and generality of a software programmable soft GPU while achieving high performance due to specialization in the persistent deep learning domain. We also propose a easy method to specialize for different domains. PDL-FGPU includes new instructions, along with micro-architecture and compiler enhancements. We evaluate both the FGPU baseline and the proposed PDL-FGPU on a modern high-end Intel Stratix 10 2800 FPGA running a set of persistent DL applications (RNN, GRU, LSTM), as well as general non-DL applications to demonstrate generality. PDL-FGPU requires 1.5-3x more ALMs, 4.4-6.4x more M20ks, and 4.6-10x more DSPs than the FGPU baseline, but improves performance by 55-727x for persistent DL applications with an average 15% degradation on general non-PDL applications. We also demonstrate that the PDL-FGPU is only 4-7x slower than the Nvidia Volta V100 GPU.

Training of convolutional neural networks (CNNs) on embedded platforms to support on-device learning is earning vital importance in recent days. Designing flexible training hardware is much more challenging than inference hardware, due to design complexity and large computation/memory requirement. In this work, we present an automatic compiler based FPGA accelerator with 16-bit fixed-point precision for complete CNN training, including Forward Pass (FP), Backward Pass (BP) and Weight Update (WU). We implemented an optimized RTL library to perform training-specific tasks and developed an RTL compiler to automatically generate FPGA-synthesizable RTL based on user-defined constraints. We present a new cyclic weight storage/access scheme for on-chip BRAM and off-chip DRAM to efficiently implement non-transpose and transpose operations during FP and BP phases, respectively. Representative CNNs for CIFAR-10 dataset are implemented and trained on Intel Stratix 10 GX FPGA using proposed hardware architecture, demonstrating up to 479 GOPS performance.

The continued rapid growth of data, along with advances in Artificial Intelligence (AI) to extract knowledge from such data, is reshaping the computing ecosystem landscape. With AI becoming an essential part of almost every end-user application, our current computing platforms are facing several challenges. The data-intensive nature of current AI models requires minimizing data movement. Furthermore, interactive intelligent datacenter-scale services require scalable and real-time solutions to provide a compelling user experience. Finally, algorithmic innovations in AI demand a flexible and programmable computing platform that can keep up with this rapidly changing field. We believe that these trends and their accompanying challenges present tremendous opportunities for FPGAs. FPGAs are a natural substrate to provide a programmable, near-data, real-time, and scalable platform for AI analytics. FPGAs are already embedded in several places where data flows throughout the computing ecosystem (e.g., "smart" network/storage, near image/audio sensors). Intel FPGAs are System-in-Package (SiP), scalable with 2.5D chiplets. They are also scalable at datacenter-scale as reconfigurable cloud, enabling real-time AI services. Using overlays, FPGAs can be programmed through software without needing long-running RTL synthesis. With further innovations, and leveraging their existing strengths, FPGAs can leap forward to realize their true potentials in AI analytics. In this talk, we first discuss the current trends in AI and big data. We then present trends in FPGA and opportunities for FPGAs in the era of AI and big data. Finally, we highlight selected research efforts to seize some of these opportunities: (1) 2.5D SiP integration of FPGA and AI chiplets to improve the performance and efficiency of AI workloads, and (2) AI overlay for FPGA to facilitate software-level programmability and compilation-speed.

Convolutional neural networks (CNN) have been shown to maintain reasonable classification accuracy when quantized to lower precisions, however, quantizing to sub 8-bit activations and weights can result in classification accuracy falling below an acceptable threshold. Techniques exist for closing the accuracy gap of limited numeric precision networks typically by means of increasing computation. This results in a trade-off between throughput and accuracy and can be tailored for different networks through various combinations of activation and weight data widths. Customizable hardware architectures like FPGAs provide the opportunity for data width specific computation through unique logic configurations leading to highly optimized processing that is unattainable by full precision networks. Specifically, ternary and binary weighted networks offer an efficient method of inference for 2-bit and 1-bit data respectively. Most hardware architectures can take advantage of the memory storage and bandwidth savings that come along with a smaller datapath, but very few architectures can take full advantage of limited numeric precision at the computation level. In this paper, we present a hardware design for FPGAs that takes advantage of the bandwidth, memory, power, and computation savings of limited numerical precision data. We provide insights into the trade-offs between throughput and accuracy for various networks and how they map to our framework. Further, we show how limited numeric precision computation can be efficiently mapped onto FPGAs for both ternary and binary cases. Starting with Arria 10, we show a 2-bit activation and ternary weighted AlexNet running in hardware that achieves 3,700 images per second on the ImageNet dataset with a top-1 accuracy of 0.49. Using a hardware modeler designed for our low numeric precision framework we project performance most notably for a 55.5 TOPS Stratix 10 device running a modified ResNet-34 with only 3.7% accuracy degradation compared with single precision.

Convolutional neural networks have been shown to maintain reasonable classification accuracy when quantized down to 8-bits, however, quantizing to sub 8-bit activations and weights can result in classification accuracy falling below an acceptable threshold. Techniques exist for increasing accuracy of sub 8-bit networks typically by means of increasing computation resulting in a trade-off between throughput and accuracy and can be tailored for different networks through combinations of activation and weight precisions. Customizable hardware architectures like FPGAs provide opportunity for data width specific computation through unique logic configurations leading to highly optimized processing that is unattainable by full precision networks. Specifically, ternary and binary weighted networks offer an efficient method of inference for 2-bit and 1-bit data respectively. In this paper, we present a hardware design for FPGAs that takes advantage of the bandwidth, memory, and computation savings of limited numerical precision data. We provide insights into the trade-offs between throughput and accuracy for various networks and how they map to our framework. Further, we show how limited numeric precision computation can be efficiently mapped onto FPGAs for both ternary and binary cases. Starting with Arria 10, we show a 2-bit activation and ternary weighted AlexNet running in hardware that achieves 3,700 images per second on the ImageNet dataset with a top-1 accuracy of 0.49. Using a hardware modeler designed for our low numeric precision framework we project performance most notably for a 55.5 TOPS Stratix 10 device running a modified ResNet-34 with only 3.7% accuracy degradation compared with single precision.

General Matrix to Matrix multiplication (GEMM) is the cornerstone for a wide gamut of applications in high performance computing (HPC), scientific computing (SC) and more recently, deep learning. In this work, we present a customizable matrix multiplication framework for the Intel HARPv2 CPU+FPGA platform that includes support for both traditional single precision floating point and reduced precision workloads. Our framework supports arbitrary size GEMMs and consists of two parts: (1) a simple application programming interface (API) for easy configuration and integration into existing software and (2) a highly customizable hardware template. The API provides both compile and runtime options for controlling key aspects of the hardware template including dynamic precision switching; interleaving and block size control; and fused deep learning specific operations. The framework currently supports single precision floating point (FP32), 16, 8, 4 and 2 bit Integer and Fixed Point (INT16, INT8, INT4, INT2) and more exotic data types for deep learning workloads: INT16xTernary, INT8xTernary, BinaryxBinary.

FPGAs or ASICs? There is a long-running debate on this. FPGAs are extremely flexible while ASICs offer top efficiency but inflexible. We believe that FPGAs and ASICs are better together, to offer both flexible and efficient solutions. We propose single-package heterogeneous 2.5D integration of FPGAs and ASICs, using Intel's Embedded Multi-Die Interconnect Bridge (EMIB). Since the ASICs are separate chips from the FPGA, this approach (1) does not change FPGA fabric, allowing re-use of existing ecosystems (FPGA chips/packaging/boards/software/etc), and (2) allows freedom in the ASIC design (area/freq/process/etc unconstrained by FPGA fabric). This approach is more effective than developing traditional stand-alone ASICs. Intel® Stratix® 10 FPGAs already have EMIBs, enabling single-package integration with other chips, or "tiles". We propose leveraging them to mix-and-match any domain-specific ASICs with Stratix10 FPGAs.

FPGAs or ASICs? FPGAs are extremely flexible while ASICs offer top efficiency. We believe that FPGAs and ASICs are better together, to offer flexibility and efficiency. We propose single-package heterogeneous 2.5D integration of FPGAs and ASICs, using Intel's Embedded Multi-Die Interconnect Bridge (EMIB). Since the ASICs are separate chips from the FPGA, this approach (1) does not change FPGA fabric, allowing re-use of existing ecosystems (FPGA chips, packaging, boards, software), and (2) allows freedom in ASIC design (area/freq/process/etc unconstrained by FPGA fabric). Intel® Stratix® 10 FPGAs already have EMIBs, enabling single-package integration with other chips, or “tiles”. We propose leveraging them to mix-and-match any domain-specific ASICs with Stratix10 FPGAs. In particular, this work focuses on deep learning (DL) domain, which demands efficient tensor (matrix/vector) operations. We propose TensorTile ASICs for Stratix10 FPGAs to provide ASIC-level tensor performance, while relying on FPGA's flexibility for application-specific operations (e.g., Winograd). Our evaluation shows: (1) a small TensorTile offer much better tensor throughput than a large Stratix102800 FPGA; (2) FPGAs and TensorTiles mix-and-match provide scalable solutions (e.g., ~69 peak INT8 TOPs with 1×TensorTile+small Stratix10-400 FPGA, to ~194 peak FP16 TOPs with 6×TensorTiles+large Stratix10-2800); (3) AlexNet performance (performance/Watt) of Intel's DL FPGA design improved by 4× (3.3×) when enhanced with 2×TensorTiles.

Intel® Stratix® 10 FPGAs offer a novel architectural feature called HyperFlex that enables an extreme degree of pipelining resulting in up to 1GHz clock frequencies. Prior work has evaluated HyperFlex on pre-production Stratix 10 FPGAs using internal designs not accessible to the general public. This paper presents an updated evaluation of HyperFlex on the latest publicly-available production Stratix 10 FPGA using open-source benchmarks. In particular, our evaluation started with seven RTL designs from existing open-source projects, carefully chosen to capture a variety of architectures (simple pipeline to pipeline with loop/M20Ks/DSPs) implementing well-known functions such as crypto, math, and image processing. An FPGA developer who was not an expert in HyperFlex then spent around 250 engineering hours to develop 24 optimized versions of these designs, following the Intel Stratix 10 FPGA HyperFlex optimization guide. Those optimized designs run at 400MHz to 850MHz. In this paper, we describe the optimizations, efforts, and results. Upon publication, those optimized designs will be open-sourced and published in GitHub.

Text analytics applications using machine learning techniques have grown in importance with ever increasing amount of data being generated from web-scale applications, social media and digital repositories. Apart from being large in size, these generated data are often unstructured and are heavily sparse in nature. The performance of these applications on current systems is hampered by hard to predict branches and low compute-per-byte ratio. This paper proposes a set of fine-grained accelerators that improve the performance and energy-envelope of these applications by an order of magnitude.

Convolutional neural networks (CNNs) are deployed in a wide range of image recognition, scene segmentation and object detection applications. Achieving state of the art accuracy in CNNs often results in large models and complex topologies that require significant compute resources to complete in a timely manner. Binarised neural networks (BNNs) have been proposed as an optimised variant of CNNs, which constrain the weights and activations to +1 or -1 and thus offer compact models and lower computational complexity per operation. This paper presents a high performance BNN accelerator on the Intel®Xeon+FPGA™ platform. The proposed accelerator is designed to take advantage of the Xeon+FPGA system in a way that a specialised FPGA architecture can be targeted for the most compute intensive parts of the BNN whilst other parts of the topology can be handled by the Xeon™ CPU. The implementation is evaluated by comparing the raw compute performance and energy efficiency for key layers in standard CNN topologies against an Nvidia Titan X Pascal GPU and other published FPGA BNN accelerators. The results show that our single-package integrated Arria™ 10 FPGA accelerator coupled with a high-end Xeon CPU can offer comparable performance and better energy efficiency than a high-end discrete Titan X GPU card. In addition, our solution delivers the best performance compared to previous BNN FPGA implementations.

Deep neural networks (DNNs) have gained popularity for their state-of-the-art accuracy and relative ease of use. DNNs rely on a growing variety of matrix multiply operations (i.e., dense to sparse, FP32 to N-bit). We propose an OpenCL-based matrix multiply design template, which enables automated design exploration to generate optimized FPGA matrix accelerators for DNN applications. Given the desired matrix operations (e.g., sparsity, data types), our template rapidly produces performance and area estimates for a variety of design variants and/or FPGA platforms. Upon identifying compelling design points and target platforms, FPGA implementations can then be generated automatically using the Intel® OpenCL™ FPGA SDK. We show the effectiveness of the template with a comparison to hand-tuned RTL, a design space exploration, and a DNN case study.

We explore techniques to significantly improve the compute efficiency and performance of Deep Convolution Networks without impacting their accuracy. To improve the compute efficiency, we focus on achieving high accuracy with extremely low-precision (2-bit) weight networks, and to accelerate the execution time, we aggressively skip operations on zero-values. We achieve the highest reported accuracy of 76.6% Top-1/93% Top-5 on the Imagenet object classification challenge with low-precision network while reducing the compute requirement by ~3× compared to a full-precision network that achieves similar accuracy. Furthermore, to fully exploit the benefits of our low-precision networks, we build a deep learning accelerator core, DLAC, that can achieve up to 1 TFLOP/mm2equivalent for single-precision floating-point operations (~2 TFLOP/mm2for half-precision), which is ~5× better than Linear Algebra Core [16] and ~4× better than previous deep learning accelerator proposal [8].

The advent of FPGA acceleration platforms with direct coherent access to processor memory creates an opportunity for accelerating applications with irregular parallelism governed by large in-memory pointer-based data structures. This paper uses the simple reference behavior of a linked-list traversal as a proxy to study the performance potentials of accelerating these applications on shared-memory processor-FPGA systems. The linked-list traversal is parameterized by node layout in memory, per-node data payload size, payload dependence, and traversal concurrency to capture the main performance effects of different pointer-based data structures and algorithms. The paper explores the trade-offs over a wide range of implementation options available on shared-memory processor-FPGA architectures, including using tightly-coupled processor assistance. We make observations of the key effects on currently available systems including the Xilinx Zynq, the Intel QuickAssist QPI FPGA Platform, and the Convey HC-2. The key results show: (1) the FPGA fabric is least efficient when traversing a single list with non-sequential node layout and a small payload size; (2) processor assistance can help alleviate this shortcoming; and (3) when appropriate, a fabric only approach that interleaves multiple linked list traversals is an effective way to maximize traversal performance.

Recurrent neural networks (RNNs) provide state-of-the-art accuracy for performing analytics on datasets with sequence (e.g., language model). This paper studied a state-of-the-art RNN variant, Gated Recurrent Unit (GRU). We first proposed memoization optimization to avoid 3 out of the 6 dense matrix vector multiplications (SGEMVs) that are the majority of the computation in GRU. Then, we study the opportunities to accelerate the remaining SGEMVs using FPGAs, in comparison to 14-nm ASIC, GPU, and multi-core CPU. Results show that FPGA provides superior performance/Watt over CPU and GPU because FPGA's on-chip BRAMs, hard DSPs, and reconfigurable fabric allow for efficiently extracting fine-grained parallelisms from small/medium size matrices used by GRU. Moreover, newer FPGAs with more DSPs, on-chip BRAMs, and higher frequency have the potential to narrow the FPGA-ASIC efficiency gap.

Deep neural networks (DNNs) are widely used in data analytics, since they deliver state-of-the-art accuracies. Binarized neural networks (BNNs) are recently proposed optimized variant of DNNs. BNNs constraint network weight and/or neuron value to either +1 or −1, which is representable in 1 bit. This leads to dramatic algorithm efficiency improvement, due to reduction in the memory and computational demands. This paper evaluates the opportunity to further improve the execution efficiency of BNNs through hardware acceleration. We first proposed a BNN hardware accelerator design. Then, we implemented the proposed accelerator on Aria 10 FPGA as well as 14-nm ASIC, and compared them against optimized software on Xeon server CPU, Nvidia Titan X server GPU, and Nvidia TX1 mobile GPU. Our evaluation shows that FPGA provides superior efficiency over CPU and GPU. Even though CPU and GPU offer high peak theoretical performance, they are not as efficiently utilized since BNNs rely on binarized bit-level operations that are better suited for custom hardware. Finally, even though ASIC is still more efficient, FPGA can provide orders of magnitudes in efficiency improvements over software, without having to lock into a fixed ASIC solution.

Sparse matrix vector multiplication (SpMV) is a linear algebra construct commonly found in machine learning (ML) algorithms, such as support vector machine (SVM). We profiled a popular SVM software (libSVM) on an energy-efficient microserver and a high-performance server for real-world ML datasets, and observed that SpMV dominates runtime. We propose a novel SpMV algorithm tailored for ML and a hardware accelerator architecture design based on this algorithm. Our evaluations show that the proposed algorithm and hardware accelerator achieves significant efficiency improvements over the conventional SpMV algorithm used in libSVM.

Maximum a posteriori probability (MAP) inference on Markov random fields (MRF) is the basis of many computer vision applications. Sequential tree-reweighted belief propagation (TRW-S) has been shown to provide very good inference quality and strong convergence properties. However, software TRW-S solvers are slow due to the algorithm's high computational requirements. A state-of-the-art FPGA implementation has been developed recently, which delivers substantial speedup over software. In this paper, we improve upon the TRW-S algorithm by using a multi-level hierarchical MRF formulation. We demonstrate the benefits of Hierarchical-TRW-S over TRW-S, and incorporate the proposed improvements on a Convey HC-1 CPU-FPGA hybrid platform. Results using four Middlebury stereo vision benchmarks show a 21% to 53% reduction in inference time compared with the state-of-the-art TRW-S FPGA implementation. To the best of our knowledge, this is the fastest hardware implementation of TRW-S reported so far.

Vertex-centric graph computations are widely used in many machine learning and data mining applications that operate on graph data structures. This paper presents GraphGen, a vertex-centric framework that targets FPGA for hardware acceleration of graph computations. GraphGen accepts a vertex-centric graph specification and automatically compiles it onto an application-specific synthesized graph processor and memory system for the target FPGA platform. We report design case studies using GraphGen to implement stereo matching and handwriting recognition graph applications on Terasic DE4 and Xilinx ML605 FPGA boards. Results show up to 14.6× and 2.9× speedups over software on Intel Core i7 CPU for the two applications, respectively.

Many promising embedded computer vision applications, such as stereo estimation, rely on inference computation on Markov Random Fields (MRFs). Sequential Tree-Reweighted Message passing (TRW-S) is a superior MRF solving method, which provides better convergence and energy than others (e.g., belief propagation). Since software TRW-S solvers are slow, custom TRW-S hardware has been proposed to improve execution efficiency. This paper proposes hardware mechanisms to further optimize TRW-S hardware efficiency, by tracking differences in input message values (residues) and skipping computation when values no longer change (residue is zero). Evaluations of our hardware mechanisms using Middlebury benchmark show 1.6x to 6x potential reduction in computation (depending on design parameters) while increasing energy by only 0.4% to 4.8%.

This paper presents a transactional specification framework (T-spec) for describing a datapath and the tool T-piper to synthesize automatically an in-order pipelined implementation with arbitrary user-specified pipeline-stage boundaries. T-spec abstractly views a datapath as executing one transaction at a time, computing the next system states based on the current ones. The synthesized pipeline maintains this semantics, yet allows concurrent execution of multiple overlapped transactions in different pipeline stages, where each stage performs a part of the next-state computation of each transaction. T-spec makes the state reading and writing events in a datapath explicit to enable T-piper to perform exact read-after-write (RAW) hazard analysis between the overlapped transactions. T-piper can automatically generate the pipeline control not only to ensure the correctness of the pipelined executions but also to minimize (using forwarding and speculation) the performance loss due to pipeline stalls in the presence of RAW dependencies. This paper reports design case studies applying T-spec and T-piper to reduced instruction set computing and complex instruction set computing processor pipeline development. In the latter, we report the results from a rapid design space exploration of 60 generated x86-subset pipelines, varying in pipeline depth, forwarding, and speculative execution, all starting from a single T-spec.

We present a technique to automatically synthesize a multithreaded in-order pipeline from a high-level unpipelined datapath specification. This work extends the previously proposed transactional specification (T-spec) and synthesis technology (T-piper). The technique not only works with instruction processors but also flexible enough to accept any sequential datapath. It maintains previously proposed non-threaded pipeline features and is enhanced with multithreading features. We report a design space exploration study of 32 multithreaded x86 processor pipelines, all synthesized from a single T-spec.

We present a transactional datapath specification (T-spec) and the tool (T-piper) to synthesize automatically an in-order pipelined implementation from it. T-spec abstractly views a datapath as executing one transaction at a time, computing next system states based on current ones. From a T-spec, T-piper can synthesize a pipelined implementation that preserves original transaction semantics, while allowing simultaneous execution of multiple overlapped transactions across pipeline stages. T-piper not only ensures the correctness of pipelined executions, but can also employ forwarding and speculation to minimize performance loss due to data dependencies. Design case studies on RISC and CISC processor pipeline development are reported.

Functional full-system simulators are powerful and versatile research tools for accelerating architectural exploration and advanced software development. Their main shortcoming is limited throughput when simulating large multiprocessor systems with hundreds or thousands of processors or when instrumentation is introduced. We propose the ProtoFlex simulation architecture, which uses FPGAs to accelerate full-system multiprocessor simulation and to facilitate high-performance instrumentation. Prior FPGA approaches that prototype a complete system in hardware are either too complex when scaling to large-scale configurations or require significant effort to provide full-system support. In contrast, ProtoFlex virtualizes the execution of many logical processors onto a consolidated number of multiple-context execution engines on the FPGA. Through virtualization, the number of engines can be judiciously scaled, as needed, to deliver on necessary simulation performance at a large savings in complexity. Further, to achieve low-complexity full-system support, a hybrid simulation technique called transplanting allows implementing in the FPGA only the frequently encountered behaviors, while a software simulator preserves the abstraction of a complete system. We have created a first instance of the ProtoFlex simulation architecture, which is an FPGA-based, full-system functional simulator for a 16-way UltraSPARC III symmetric multiprocessor server, hosted on a single Xilinx Virtex-II XCV2P70 FPGA. On average, the simulator achieves a 38x speedup (and as high as 49×) over comparable software simulation across a suite of applications, including OLTP on a commercial database server. We also demonstrate the advantages of minimal-overhead FPGA-accelerated instrumentation through a CMP cache simulation technique that runs orders-of-magnitude faster than software.

Functional full-system simulators are powerful and versatile research tools for accelerating architectural exploration and advanced software development. Their main shortcoming is limited throughput when simulating systems with hundreds of processors or more. To overcome this bottleneck, we propose the PROTOFLEX simulation architecture, which uses FPGAs to accelerate simulation. Prior FPGA approaches that prototype a complete system in hardware are either too complex when scaling to large-scale configurations or require significant effort to provide full-system support. In contrast, PROTOFLEX reduces complexity by virtualizing the execution of many logical processors onto a consolidated set of multiple-context execution engines on the FPGA. Through virtualization, the number of engines can be judiciously scaled, as needed, to deliver on necessary simulation performance. To achieve low-complexity full-system support, a hybrid simulation technique called transplanting allows implementing in the FPGA only the frequently encountered behaviors, while a software simulator preserves the abstraction of a complete system

This paper presents the active cache emulator (ACE), a novel field-programmable gate-array (FPGA)-based emulator that models an L3 cache actively and in real-time. ACE leverages interactions with its host system to model the target system. Unlike most existing FPGA-based cache emulators that collect only memory traces from their host system, ACE provides feedback to its host by injecting delays to time dilate the host system such that it experiences hit/miss latencies of the emulated cache. Such active emulation expands the context of performance evaluations by allowing measurements of system performance metrics (e.g., CPI, operations per second, frame rate) in addition to the typical cache-specific performance metrics (e.g., miss ratio) provided by existing emulators. ACE is designed to interface with a front-side bus (FSB) of a typical Pentium-based PC system. ACE utilizes the FSB snoop stall mechanism to inject delays into the system. At present, ACE is implemented using a Xilinx XC2V6000 FPGA running at 66 MHz, the same speed as its host's FSB. Verification of ACE includes using the cache calibrator and RightMark memory analyzer software to confirm proper detection of the emulated cache by the host system, and comparing ACE results with SimpleScalar software simulations. Finally, ACE is used to study L3 caches for compute-intensive, throughput-oriented, and real-time gaming benchmarks (SPEC-CPU2000, SPEC-JBB2000, Quake3). The study shows that analyzing only cache-specific metrics, as done by existing L3 cache studies with FPGA emulators, is insufficient. Active emulation mitigates this issue by providing a broader performance view, allowing researchers make better research conclusion.

PROTOFLEX is an FPGA-accelerated hybrid simulation/emulation platform designed to support large-scale multiprocessor hardware and software research. Unlike prior attempts at FPGA multiprocessor system emulators, PROTOFLEx emulates full-system fidelityi.e., runs stock commercial operating systems with I/O support. This is accomplished without undue effort by leveraging a hybrid emulation technique called transplanting. Our transplant technology uses FPGAs to accelerate only common-case behaviors while relegating infrequent, complex behaviors (e.g., I/O devices) to software simulation. By working in concert with existing full-system simulators, transplanting avoids the costly and unnecessary construction of the entire target system in FPGA. We report preliminary findings from a working hybrid PROTOFLEx emulator of an UltraSPARC workstation running Solaris 8. We have also started developing a novel multiprocessor emulation approach that interleaves the execution of many (10s to 100s) processor contexts onto a shared emulation engine. This approach decouples the scale and complexity of the FPGA hostfrom the simulated system size but neverthe-less enables us to scale the desired emulation performance by the number of emulation engines used. Together, the transplant and interleaving techniques will enable us to develop full-system FPGA emulators of up to thousands of processors without an overwhelming development effort.

This paper presents the design, implementation, and verification of the Active Cache Emulator (ACE), a novel FPGA-based emulator that models an L3 cache actively and in real-time. ACE leverages interactions with its host system to model the target system (i.e. hypothetical system under study). Unlike most existing FPGA-based cache emulators that collect only memory traces from their host system, ACE provides feedback to its host by modeling the impact of the emulated cache on the system. Specifically, delays are injected to time dilate the host system which then experiences hit/miss latencies of the emulated cache. Such active emulation expands the context of performance measurements by capturing processor performance metrics (e.g. cycle per instruction) in addition to measuring the typical cache-specific performance metrics (e.g. miss ratio).ACE is designed to interface with a front-side bus (FSB) of a typical Pentium®-based PC system. To actively emulate cache latencies, ACE utilizes the snoop stall mechanism of the FSB to inject delays to the system. At present, ACE is implemented using a Xilinx XC2V6000 FPGA running at 66MHz, the same speed as its host's FSB. Verification of ACE includes using the Cache Calibrator and RightMark Memory Analyzer software to confirm proper detection of the emulated cache by the host system, and comparing ACE results with SimpleScalar software simulations.

With the proliferation of e-businesses, Java™ Middleware and OLTP applications are gaining importance. As the gap between CPU and memory latencies continues to increase, the performance of these applications running on multiprocessor systems will become further limited by the memory system. This study characterizes the memory behavior of such applications using the SPECjAppServer2002 and TPC-C benchmarks running on a real multiprocessor system. More specifically, the shared and private L3 caches with invalidation- and update-based coherence protocols are evaluated using the Programmable Hardware-Assisted Cache Emulator (PHA$E). We found that coherency misses increase with larger private L3 caches, constituting up to more than 15% of all misses for both benchmarks. Additionally, a saturation point was observed at which employing larger private cache yields no further improvement in miss ratio. Conversely, the shared L3 cache design was observed to be more scalable since it does not suffer from coherence misses. Our limit study shows that the existing Write-Broadcast policy, which updates line copies in other caches during a write on a shared line, has the potential to simultaneously reduce private cache miss ratio and bus traffic. For example, at 64MB, it reduces the miss ratio by 53% and 44% respectively for SPECjAppServer2002 and TPC-C, while lowering the bus traffic by 18% and 11%. In overall, the policy can eliminate the aforementioned saturation point and allows private cache miss ratio that is comparable with the miss ratio of a shared cache.