A key challenge in scaling shared-L1 multi-core clusters towards many-core (more than 16 cores) configurations is to ensure low-latency and efficient access to the L1 memory. In this work we demonstrate that it is possible to scale up the shared-L1 architecture: We present MemPool, a 32 bit many-core system with 256 fast RV32IMA “Snitch” cores featuring application-tunable execution units, running at 700 MHz in typical conditions (TT/0.80 V/25 °C). MemPool is easy to program, with all the cores sharing a global view of a large L1 scratchpad memory pool, accessible within at most 5 cycles. In MemPool's physical-aware design, we emphasized the exploration, design, and optimization of the low-latency processor-to-L1-memory interconnect. We compare three candidate topologies, analyzing them in terms of latency, throughput, and back-end feasibility. The chosen topology keeps the average latency at fewer than 6 cycles, even for a heavy injected load of 0.33 request/core/cycle. We also propose a lightweight addressing scheme that maps each core private data to a memory bank accessible within one cycle, which leads to performance gains of up to 20 % in real-world signal processing benchmarks. The addressing scheme is also highly efficient in terms of energy consumption since requests to local banks consume only half of the energy required to access remote banks. Our design achieves competitive performance with respect to an ideal, non-implementable full-crossbar baseline.

Electronic-skin aiming to mimic human skin is becoming a reality and systems able to process data close to the sensors are required to reduce latency and power consumption. This paper presents the design and implementation of an energy efficient smart system for tactile sensing based on a RISC-V parallel ultra-low power platform (PULP). The PULP processor, called Mr. Wolf, performs the on-board classification of different touch modalities. This demonstrates the promising use of on-board classification for emerging robot and prosthetic applications. Experimental results demonstrate the effectiveness of the platform on improving the energy efficiency of the online classification. In our experiments, Mr. Wolf runs 3.6 times faster than an ARM Cortex M4F (STM32F40), consuming only 28 mW. The proposed platform achieves 15× better energy efficiency, than the classification done on the STM32F40, consuming only 81mJ per classification.

Current ultra-low power smart sensing edge devices, operating for years on small batteries, are limited to low-bandwidth sensors, such as temperature or pressure. Enabling the next generation of edge devices to process data from richer sensors such as image, video, audio, or multi-axial motion/vibration has huge application potential. However, edge processing of data-rich sensors poses the extreme challenge of squeezing the computational requirements of advanced, machine-Iearning-based near-sensor data analysis algorithms (such as Convolutional Neural Networks) within the mW-range power envelope of always-ON battery-powered IoT end-nodes. To address this challenge, we propose GAP-8: a multi-GOPS fully programmable RISC-V IoT-edge computing engine, featuring a 8-core cluster with CNN accelerator, coupled with an ultra-low power MCU with 30 μW state-retentive sleep power. GAP-8 delivers up to 10 GMAC/s for CNN inference (90 MHz, 1.0V) at the energy efficiency of 600 GMAC/s/W within a worst-case power envelope of75 mW.

Energy efficiency is a crucial aspect in modern SoCs. Common strategies like aggressive voltage scaling and parallel processing have enabled major improvements in active energy efficiency. However, the big impact of process variations, as well as the temperature sensitivity of devices operating in near threshold force digital designers to adopt very conservative margins for timing closure.

Endpoint devices for Internet-of-Things not only need to work under extremely tight power envelope of a few milliwatts, but also need to be flexible in their computing capabilities, from a few kOPS to GOPS. Near-threshold (NT) operation can achieve higher energy efficiency, and the performance scalability can be gained through parallelism. In this paper, we describe the design of an open-source RISC-V processor core specifically designed for NT operation in tightly coupled multicore clusters. We introduce instruction extensions and microarchitectural optimizations to increase the computational density and to minimize the pressure toward the shared-memory hierarchy. For typical data-intensive sensor processing workloads, the proposed core is, on average, 3.5× faster and 3.2× more energy efficient, thanks to a smart L0 buffer to reduce cache access contentions and support for compressed instructions. Single Instruction Multiple Data extensions, such as dot products, and a built-in L0 storage further reduce the shared-memory accesses by 8× reducing contentions by 3.2×. With four NT-optimized cores, the cluster is operational from 0.6 to 1.2 V, achieving a peak efficiency of 67 MOPS/mW in a low-cost 65-nm bulk CMOS technology. In a low-power 28-nm FD-SOI process, a peak efficiency of 193 MOPS/mW (40 MHz and 1 mW) can be achieved.

Computer vision (CV) based on Convolutional Neural Networks (CNN) is a rapidly developing field thanks to CNN's flexibility, strong generalization capability and classification accuracy (matching and sometimes exceeding human performance). CNN-based classifiers are typically deployed on servers or high-end embedded platforms. However, their ability to “compress” low information density data such as images into highly informative classification tags makes them extremely interesting for wearable and IoT scenarios, should it be possible to fit their computational requirements within deeply embedded devices such as visual sensor nodes. We propose a 65nm system-on-chip implementing a hybrid HW/SW CNN accelerator while meeting this energy efficiency target. The SoC integrates a near-threshold parallel processor cluster [1] and a hardware accelerator for convolution-accumulation operations [2], which constitute the basic kernel of CNNs: it achieves peak performance of 11.2 GMAC/s @ 1.2 V and peak energy efficiency of 261 GMAC/s/W @ 0.65V.

Limited power budgets and the need for high performance computing have led to platform customization with a number of accelerators integrated with CMPs. In order to study customized architectures, we model four customization design points and compare their performance and energy across a number of computer vision workloads. We analyze the limitations of generic architectures and quantify the costs of increasing customization using these micro-architectural design points. This analysis leads us to develop a framework consisting of low-power multi-cores and an array of configurable micro-accelerator functional units. Using this platform, we illustrate dataflow and control processing optimizations that provide for performance gains similar to custom ASICs for a wide range of vision benchmarks.

Baseline RISC instruction sets for ultra-low power processors are constantly being tuned to reduce cycle count when executing computation-intensive applications. Performance improvements often come at a non-negligible price in terms of area and critical path length and imply deeper pipelines and complex memory interfaces. This penalizes control-intensive code execution and significantly increases cost and complexity of building multi-core clusters. In addition, some extensions are not easily exploited by compilers and may increase code development effort, especially when considering parallel applications. In this paper we describe our efforts in enhancing a baseline open ISA (OpenRISC) and its LLVM compiler back-end to significantly reduce execution cycles while minimizing the impact on core micro-architecture complexity, number of pipeline stages, area and power. In addition, we improved the core micro-architecture to streamline its integration in a tightly-coupled cluster, sharing instruction cache and data memory, thereby further enhancing parallel execution efficiency. The combined effect of ISA, compiler and micro-architecture evolution gives an average energy efficiency boost of 59% on vector intensive code and 41% otherwise, at an area and power increase of 2.3% and 18% on a four-core processor cluster.

Research on Networks on Chips (NoCs) has spanned over a decade and its results are now visible in some products. Thus the seminal idea of using networking technology to address the chip-level interconnect problem has been shown to be correct. Moreover, as technology scales down in geometry and chips scale up in complexity, NoCs become the essential element to achieve the desired levels of performance and quality of service while curbing power consumption levels. Design and timing closure can only be achieved by a sophisticated set of tools that address NoC synthesis, optimization and validation.

Networks-on-chip have been developed in the last few years to address the scalability challenges of global on-chip communication. VLSI technology is now rapidly moving into vertical stacking to overcome fundamental communication and integration bottlenecks, however this technology is not mature yet, and significant reliability challenges must be overcome. In this paper we describe our effort in establishing a 3DNoC design flow and in designing circuits and architectural solutions for variability and reliability characterization and tolerance.

Near-Threshold Circuits achieve ultra-low energy operating with significant performance improvement and noise immunity as compared to sub-threshold circuits. However, near-threshold circuit performance is highly sensitive to static and dynamic threshold voltage variations. This makes designing circuits for a target performance very difficult, and post-silicon tunability is required to achieve performance targets without taking huge design margins. In this work, we tackle this problem by proposing a novel dual-Vdd technique for near-threshold operation and show that one can tune the performance of a circuit in a fine-grained manner by powering an optimal sub-set of rows with a slightly higher supply voltage than the rest, without incurring the large cost of distributed level shifters. By varying the percentage of rows at a slightly higher voltage, one can trade-off performance and power in a fine-grained manner. Experimental results show that by employing our dual-Vdd technique, we can improve the performance of several benchmarks up-to 45% while achieving more than 50% lower power as compared to single-Vdd implementations.

The quest for technologies with superior device characteristics has showcased Carbon Nanotube Field Effect Transistors (CNFETs) into limelight. Among the several design aspects necessary for today's grail in CNFET technology, achieving functional immunity to Carbon Nanotube (CNT) manufacturing issues (such as mispositioned CNTs and metallic CNTs) is of paramount importance. In this work we present a new design technique to build compact layouts while ensuring 100% functional immunity to mispositioned CNTs. Then, as second contribution of this work, we have developed a CNFET Design Kit (DK) to realize a complete design flow from logic-to-GDSII traversing the conventional CMOS design flow. This flow enables a framework that allows accurate comparison between CMOS and CNFET-based circuits. This paper also presents simulation results to illustrate such analysis, namely, a CNFET-based inverter can achieve gains, with respect to the Energy-Delay Product (EDP) metric, of more than 4× in delay, 2× in energy/cycle and significant area savings (more than 30%) when compared to a corresponding CMOS inverter benchmarked with an industrial 65nm technology.

Nanometer CMOS scaling has resulted in greatly increased circuit variability, with extremely adverse consequences on design predictability and yield. A number of recent works have focused on adaptive post-fabrication tuning approaches to mitigate this problem. Adaptive Body Bias (ABB) is one of the most successful tuning ldquoknobsrdquo in use today in high-performance custom design. Through forward body bias (FBB), the threshold voltage of the CMOS devices can be reduced after fabrication to bring the slow dies back to within the range of acceptable specs. FBB is usually applied with a very coarse core-level granularity at the price of a significantly increased leakage power. In this paper, we propose a novel, physically clustered FBB scheme on row-based standard-cell layout style that enables selective forward body biasing of only of the rows that contain most timing critical gates, thereby reducing leakage power overhead. We propose exact and heuristic algorithms to partition the design and allocate optimal body bias voltages to achieve minimum leakage power overhead. This style is fully compatible with state-of-the-art commercial physical design flows and imposes minimal area blowup. Benchmark results show large leakage power savings with a maximum savings of 30% in case of 5% compensation and 47.6% in case of 10% compensation with respect to block-level FBB and minimal implementation area overhead.

Leakage power is a serious concern in nanometer CMOS technologies. In this paper we focus on leakage reduction through automatic insertion of sleep transistors for power gating in standard cell based designs. In particular, we propose clustering algorithms for row- based power-gating methodology which is based on using rows of the layout as the granularity for clustering. Our clustering methodology does timing and area constraint driven power-gating in contrast to only timing driven power-gating as proposed in the previous works. We present two distinct clustering algorithms with different accuracy-efficiency trade-off. An optimal one, which exploits a 0-1 or binary integer programming approach, and a heuristic one, which resorts to an implicit enumeration of the layout rows. Results show that, for all the benchmarks, the leakage power savings, as compared to previous techniques, are more than 75% when we have the same timing constraints but half sleep transistor area and at least 60% when area constraint is set at one fourth. We also show that we can perform clustering with no speed degradation and achieve maximum leakage power savings up-to 83%.

Leakage power reduction in nano-CMOS designs has gained tremendous interest both in academia and industry. Many techniques have been proposed in the literature for leakage power reduction and one of the prominent techniques for leakage power reduction is the use of sleep transistors as power-gating elements to cut-off sub-threshold leakage current in circuits when they are in stand-by mode. Although sleep transistor insertion is very effective in cutting-off leakage, it also incurs timing, area and routing overhead. Since most of the sleep transistor insertion methodologies do post layout insertion, care should be taken such that there is minimal perturbation of the original layout. Over design of sleep transistors cells and sub-optimal sleep transistor placement must be avoided to achieve final design closure. Since the sleep transistor area plays an important and prominent role in this aspect, it necessitates for optimal sleep transistor sizing and synthesis technique under area constraints. In this paper, we first provide a methodology for optimal sleep transistor synthesis under given area constraints. We then apply our technique to the general timing and area constraint driven row-based power-gating methodology proposed in [13] and show how optimal low leakage designs with constraints on timing and area can be designed

Power-gating has proved to be one of the most effective solutions for reducing stand-by leakage power in nanometer-scale CMOS circuits, and different strategies and algorithms for its application have been proposed recently. Unfortunately, power- gating comes with its own set of costs: Performance degradation, area increase, dynamic power increase and routing congestion. When a decision to power-gate a design has to be taken, pros and cons of power-gating have to be properly weighted to achieve optimal results. In this paper, we define "Figures of Merit" (FoMs) for power-gating, which can be used by designers to better understand the benefits and costs of power-gating, thereby allowing them to achieve optimal results. We then quantify the FoMs by applying a state-of-the-art, industry-strength power- gating flow on a set of designs implemented onto an industrial 65 nm CMOS process, and provide insightful discussion on how optimum power-gating can be achieved.

Clustered sleep transistor insertion is an effective leakage power reduction technique that is well-suited for integration in an automated design flow and offers a flexible tradeoff between area, delay overhead and turn-on transition time. In this work, we focus on the design of a family of sleep transistor cells, fully compatible with the physical design rules of a commercial 65nm CMOS library. We describe circuit-level and layout optimizations, as well as the cell characterization procedure required to support automated sleep transistor cell selection and instantiation in a clustered power-gating insertion flow.

In this paper we focus on leakage reduction through automatic insertion of sleep transistors using a row-based granularity. In particular, we tackle here the two main issues involved in this methodology: (i) Clustering and (ii) the interfacing of power-gated and non power-gated regions within the same block. The clustering algorithm automatically selects an optimal subset of rows that can be power-gated with a tightly controlled delay overhead. We then address the issue of interfacing different gated regions and propose a novel technique to address this issue with minimal area and power penalty. Our approach is compatible with state-of-the art logic and physical synthesis flows and it does not significantly impact design closure. We achieve leakage power reductions as high as 89% for a set of standard benchmarks, with minimum timing and area overhead.

As embedded computing evolves towards ever more powerful architectures, the challenge of properly interconnecting large numbers of on-chip computation blocks is becoming prominent. Networks-on-chip (NoCs) have been proposed as a scalable solution to both physical design issues and increasing bandwidth demands. However, this claim has not been fully validated yet, since the design properties and tradeoffs of NoCs have not been studied in detail below the 100 nm threshold. This work is aimed at shedding light on the opportunities and challenges, both expected and unexpected, of NoC design in nanometer CMOS. We present fully working 65 nm NoC designs, a complete NoC synthesis flow and detailed scalability analysis

With technology scaling, the wire delay as a fraction of the total delay is increasing, and the communication architecture is becoming a major bottleneck for system performance in systems on chip (SoCs). A communication-centric design paradigm, networks on chip (NoCs), has been proposed recently to address the communication issues of SoCs. As the geometries of devices approach the physical limits of operation, NoCs will be susceptible to various noise sources such as crosstalk, coupling noise, process variations, etc. Designing systems under such uncertain conditions become a challenge, as it is harder to predict the timing behavior of the system. The use of conservative design methodologies that consider all possible delay variations due to the noise sources, targeting safe system operation under all conditions will result in poor system performance. An aggressive design approach that provides resilience against such timing errors is required for maximizing system performance. In this paper, we present T-error, which is a timing-error-tolerant aggressive design method to design the individual components of the NoC (such as switches, links, and network interfaces), so that the communication subsystem can be clocked at a much higher frequency than a traditional conservative design (up to 1.5x increase in frequency). The NoC is designed to tolerate timing errors that arise from overclocking without substantially affecting the latency for communication. We also present a way to dynamically configure the NoC between the overclocked mode and the normal mode, where the frequency of operation is lower than or equal to the traditional design's frequency, so that the error recovery penalty is completely hidden under normal operation. Experiments on several benchmark applications show large performance improvement (up to 33% reduction in average packet latency) for the proposed system when compared to traditional systems.