In recent years, neural network accelerators have been shown to achieve both high energy efficiency and high performance for a broad application scope within the important category of recognition and mining applications. Still, both the energy efficiency and performance of such accelerators remain limited by memory accesses. In this paper, we focus on image applications, arguably the most important category among recognition and mining applications. The neural networks which are state-of-the-art for these applications are convolutional neural networks (CNNs), and they have an important property: weights are shared among many neurons, considerably reducing the neural network memory footprint. This property allows to entirely map a CNN within an SRAM, eliminating all DRAM accesses for weights. By further hoisting this accelerator next to the image sensor, it is possible to eliminate all remaining DRAM accesses, i.e., for inputs and outputs. In this paper, we propose such a CNN accelerator, placed next to a CMOS or CCD sensor. The absence of DRAM accesses combined with a careful exploitation of the specific data access patterns within CNNs allows us to design an accelerator which is highly energy-efficient. We present a single-core implementation down to the layout at 65 nm, with a modest footprint of 5.94mm $^{\boldsymbol {2}}$ and consuming only 336mW, but still about $\boldsymbol {30\times }$ faster than high-end GPUs. For visual processing with higher resolution and frame-rate requirements, we further present a multicore implementation with elevated performance.

In recent years, neural network accelerators have been shown to achieve both high energy efficiency and high performance for a broad application scope within the important category of recognition and mining applications. Still, both the energy efficiency and performance of such accelerators remain limited by memory accesses. In this paper, we focus on image applications, arguably the most important category among recognition and mining applications. The neural networks which are state-of-the-art for these applications are Convolutional Neural Networks (CNN), and they have an important property: weights are shared among many neurons, considerably reducing the neural network memory footprint. This property allows to entirely map a CNN within an SRAM, eliminating all DRAM accesses for weights. By further hoisting this accelerator next to the image sensor, it is possible to eliminate all remaining DRAM accesses, i.e., for inputs and outputs. In this paper, we propose such a CNN accelerator, placed next to a CMOS or CCD sensor. The absence of DRAM accesses combined with a careful exploitation of the specific data access patterns within CNNs allows us to design an accelerator which is 60&times more energy efficient than the previous state-of-the-art neural network accelerator. We present a full design down to the layout at 65 nm, with a modest footprint of 4.86mm 2 and consuming only 320mW, but still about 30× faster than high-end GPUs.

In recent years, neural network accelerators have been shown to achieve both high energy efficiency and high performance for a broad application scope within the important category of recognition and mining applications. Still, both the energy efficiency and performance of such accelerators remain limited by memory accesses. In this paper, we focus on image applications, arguably the most important category among recognition and mining applications. The neural networks which are state-of-the-art for these applications are Convolutional Neural Networks (CNN), and they have an important property: weights are shared among many neurons, considerably reducing the neural network memory footprint. This property allows to entirely map a CNN within an SRAM, eliminating all DRAM accesses for weights. By further hoisting this accelerator next to the image sensor, it is possible to eliminate all remaining DRAM accesses, i.e., for inputs and outputs. In this paper, we propose such a CNN accelerator, placed next to a CMOS or CCD sensor. The absence of DRAM accesses combined with a careful exploitation of the specific data access patterns within CNNs allows us to design an accelerator which is 60× more energy efficient than the previous state-of-the-art neural network accelerator. We present a full design down to the layout at 65 nm, with a modest footprint of 4.86mm 2 and consuming only 320mW, but still about 30× faster than high-end GPUs.

This book describes a new design approach for energy-efficient, Domain-Specific Instruction set Processor (DSIP) architectures for the wireless baseband domain. The innovative techniques presented enable co-design of algorithms, architectures and technology, for efficient implementation of the most advanced technologies. To demonstrate the feasibility of the authors design approach, case studies are included for crucial functionality of advanced wireless systems with increased computational performance, flexibility and reusability. Designers using this approach will benefit from reduced development/product costs and greater scalability to future process technology nodes.

Emerging 4G wireless communication systems need to deliver much higher data rates, more flexibility, and a significantly higher energy efficiency than current systems. To cope with this immense increase of requirements, new design approaches are a necessity. This paper focuses on the design of an advanced multiple-input-multiple-output (MIMO) detector, which is typically a bottleneck in the wireless receiver. In the proposed template-based design approach innovative architecture concepts, such as very wide register and distributed loop buffer, and algorithm-architecture co-optimizations are combined. The resulting MIMO detector processor, which is scalable to eight and more antennas, achieves a high area efficiency of 571 GOPS/mm(2) and a high energy efficiency of 3.3 GOPS/mW in the Taiwan Semiconductor Manufacturing Company (TSMC) 40-nm technology. By exploiting the dynamically varying requirements, the proposal has the potential to achieve a higher average energy efficiency than an application-specific integrated circuit (ASIC) equivalent. However, a penalty in total area consumption exists. The proposed architecture style offers an interesting and a very promising tradeoff in between the traditional ASIC and the other programmable processor solutions.

The reduction of the energy consumption in the domain of the embedded systems is becoming the most important design goal due to the increasing use of battery powered consumer devices. Previous research has pointed out the instruction memory organisation as one of the major sources of energy consumption of the embedded systems. Due to this fact, the introduction of any enhancement in this component of the system becomes crucial in order to decrease this energy bottleneck. The purpose of this paper is to present a high-level energy analysis of the loop buffer schemes that exist in the embedded systems. The crucial energy analysis that is presented in this paper not only proposes a method to evaluate different loop buffer schemes for a certain application, but also guides embedded systems designers to make the correct decision in the trade-offs that exist between the energy budget, the required performance, and the area cost of the embedded system. Experimental results used in this analysis show that, the search of energy savings (up to 76%) has to take into account the performance penalty, the area cost, and the impact of the implementation technology in order to choose the most suitable enhancement that has to be introduced in the instruction memory organisation from the point of view of the energy consumption.

The use of distributed loop buffer architectures with incompatible loop-nest organisations allows the execution of incompatible loops in parallel with minimal hardware overhead. Due to this fact, the utilisation of these distributed and scalable architectures in embedded systems is a promising option to improve the energy efficiency of the instruction memory organisations that exist in these systems. This paper proposes and analyses non-overlapping and complementary implementation options for distinct partitions of the design space that is related to distributed loop buffer architectures. The high-level trade-off analysis of the proposed implementations is crucial to present the correct process design that an embedded systems designer has to follow in order to have an efficient distributed loop buffer architecture for a certain application. Results show that, with an increase of about 6.5 % in the energy consumption of the control logic that exists in the instruction memory organisation, the overall energy consumption of the instruction memory organisation can be reduced by 6 % to 22 %, when distributed loop buffer architectures with incompatible loop-nest organisations are used instead of clustered loop buffer architectures with shared loop-nest organisations architectures.

Data-parallel processing is a widely applicable technique, which can be implemented on different processor styles, with varying capabilities. Here we address single or multi-core data-parallel instruction-set processors. Often, handling and reorganisation of the parallel data may be needed because of diverse needs during the execution of the application code. Signal word-length considerations are crucial to incorporate because they influence the outcome very strongly. This paper focuses on the broader solution space of selecting sub-word lengths (at design time) including especially hybrids, so that mapping on these data parallel single/multi-core processors is more energy-efficient. Our goal is to introduce systematic exploration techniques so that part of the designers effort is removed. The methodology is evaluated on a representative application driver for a number of data-path variants and the most promising trade-off points are indicated. The range of throughput-energy ratios among the different mapping implementations is spanning a factor of 2.2.