Medical multiple-choice question answering (MCQA) benchmarks implicitly assume that large language models (LLMs) should always commit to an answer.However, in clinical practice, uncertainty is pervasive and abstaining is often the safest action.We introduce MedQAbstain, a benchmark explicitly designed to evaluate medical abstention under uncertainty.MedQAbstain repurposes standard medical MCQA datasets by removing the gold answer and introducing an explicit "I abstain" option, framed as a safety-critical decision with clinical consequences.The benchmark supports systematic analysis across abstention regimes, distractor complexity, and input modalities, and elicits self-reported model confidence to study calibration.Across all settings, we find that state-of-the-art LLMs systematically overcommit, rarely abstaining even when the question itself is hidden.These results reveal a fundamental mismatch between LLM behavior and clinical norms, highlighting abstention as a critical but overlooked dimension of medical decision-making evaluation. 1 * Equal contribution (co-first authors).

Resource-constrained embedded systems must efficiently execute heterogeneous workloads spanning multiple application domains. This paper presents a quantitative comparison between clustered SPMD multi-core execution and SIMD vector execution using a common benchmark suite representative of several embedded scenarios. Experiments are conducted on two open-source RISC-V architectures—the PULP cluster and the Spatz vector processor—combining benchmark results with post-layout measurements from GlobalFoundries 12 nm (GF12) implementations to enable a fair Performance, Power, and Area (PPA) assessment. For compute-intensive workloads with regular memory access patterns, Spatz achieves up to 6.86 × higher performance and 2.52 × higher energy efficiency than the PULP cluster, thanks to its higher operating frequency (1 GHz vs. 600 MHz) enabled by a more regular microarchitecture and predictable access patterns well matched to the applications. Conversely, workloads characterized by irregular memory accesses, complex control flow, reductions, or loop-carried dependencies favor the PULP cluster, which delivers up to 3.54 × higher performance and 5.72 × higher energy efficiency. Overall, the results show that vector execution can offer substantial performance and efficiency benefits when well matched to the application domain. However, its higher degree of specialization increases sensitivity to workload characteristics, making it a high-reward but high-risk choice as a standalone System on-Chip (SoC) accelerator that requires strong priori knowledge of the target workloads.

Named entity recognition (NER) in medicine is challenging due to specialized terminology, inconsistent annotation guidelines, and the continuous emergence of new entity types—requiring models that can adapt to unseen targets. Large language models (LLMs) exhibit strong generalization but are impractical for scalable deployment, whereas recent encoder-only approaches leverage entity names for zero-shot inference but struggle with disambiguation in complex domains. We introduce OpenBioNER-v2, a family of lightweight transformer encoders (15M–110M parameters) designed for zero-shot recognition of biomedical and clinical entities by conditioning on natural language descriptions of target types. Our cross-encoder architecture jointly models input text and entity-type descriptions, enabling semantic matches. Pretrained on LLM-generated silver annotations and multi-view descriptions covering thousands of medical types, OpenBioNER-v2 achieves state-of-the-art results across 11 benchmarks—including a new dataset for personal de-identification. Variants with ≤ 56M parameters outperform both large and small language models, such as UniversalNER and GliNER. Ablation studies reveal effective strategies for formulating descriptions. All data, code, and model checkpoints are publicly released under open-science principles.

Alessio Cocchieri, Luca Ragazzi, Paolo Italiani, Giuseppe Tagliavini, Gianluca Moro. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The TRISTAN project aims to expand and industrialize the European RISC-V ecosystem to compete effectively with existing commercial alternatives. This initiative specifically targets the critical challenges in the development of Electronic Design Automation (EDA) tools, essential for RISC-V-based solutions, by leveraging the synergy between the open-source community and industrial solutions. This paper presents an overview of the current landscape of TRISTAN's EDA flow, highlighting specific tools and methodologies that streamline the early design phases of RISC-V-based systems. We explore the unique features of these tools, emphasizing how they complement each other to strengthen the overall design process.

Recent advances in large language models (LLMs) have demonstrated strong mathematical reasoning abilities, even in visual contexts, with some models surpassing human performance on existing benchmarks.However, these benchmarks lack structured age categorization, clearly defined skill requirements, and-crucially-were not designed to assess human performance in international competitions.To address these limitations, we introduce MATHGAMES, a new benchmark of 2,183 high-quality mathematical problems (both text-only and multimodal) in an openended format, sourced from an international mathematical games championships.Spanning seven age groups and a skill-based taxonomy, MATHGAMES enables a structured evaluation of LLMs' mathematical and logical reasoning abilities.Our experiments reveal a substantial gap between state-of-theart LLMs and human participants-even 11year-olds consistently outperform some of the strongest models-highlighting the need for advancements.Further, our detailed error analysis offers valuable insights to guide future research.The data is publicly available at https:// disi-unibo-nlp.github.io/math-games/. * Equal contribution (co-first authors).Number of participants per score C1 (11-13 y/o) 1413 C2 (13-15 y/o) 709 L1 (15-18 y/o) 338 L2 (18-20 y/o) 177 GP (20-25 y/o) 112 0 10 20 30 40 50 60 70 80 90 100 Competition Score (%) HC (25+ y/o) 25 Gemini-2.0-Flash-ThinkGemini-1.5-ProGPT-4o GPT-4o-mini Gemini-1.5-FlashGemini-1.5-Flash-8BEarly Teenager (11-13 y/o) Late Teenager (13-18 y/o) Adult (18-25+ y/o) Question: How many small spheres of different colors are there in the figure?Question: In figure you see tennis balls placed on top of each other, forming at each "plane" of the squares, without holes in the middle.The highest level contains only one ball; the second, coming down, contains 4; the third contains 9 and so on.If you use 7714 balls, how many floors will your pyramid of tennis balls be constituted?Question: In figure you see a pentagonal tile, quite singular, whose sides BC and AE measure 1 dm while AB measures 2 dm.Which is in cm 2 , rounded to the nearest cm 2 , the area of our tile?(If necessary, use 1,414 for √ 2 and 1,732 for √ 3).

The early detection of structural malfunctions requires the installation of real-time monitoring systems ensuring continuous access to the damage-sensitive information; nevertheless, it can generate bottlenecks in terms of bandwidth and storage. Deploying data reduction techniques at the edge is recognized as a proficient solution to reduce the systemfs network traffic. However, the most effective solutions currently employed for the purpose are typically based on memory and power-hungry algorithms, making their embedding on resourceconstrained devices very challenging; this is the case of vibration data reduction based on System Identification (SysId) models. This paper presents PARSY-VDD, a fully optimized PArallel end-to-end software framework based on SYstem identification for Vibration-based Damage Detection, as a suitable solution to perform damage detection at the edge in a time and energyefficient manner, avoiding streaming raw data to the cloud. First, we evaluate the damage detection capabilities of PARSYVDD with two benchmarks: a bridge and a wind turbine blade, showcasing the robustness of the end-to-end approach. Then, we deploy PARSY-VDD on both commercial single-core (STM32 family) and a specific multi-core (GAP9) edge device. We introduce an architecture-agnostic algorithmic optimization for SysId, improving the execution by 90× and reducing the consumption by 85× compared with the state-of-the-art SysId implementation on GAP9. Results show that by utilizing the unique parallel computing capabilities of GAP9, the execution time is 751 μ s with the high-performance multi-core solution operating at 370MHz and 0.8V, while the energy consumption is 37 μ J with the low-power solution operating at 240MHz and 0.65V. Compared with other single-core implementations based on STM32 microcontrollers, the GAP9 high-performance configuration is 76× faster, while the low-power configuration is 360× more energy efficient.

Balancing energy efficiency and high performance in embedded systems requires fine-tuning hardware and software components to co-optimize their interaction. In this work, we address the automated optimization of memory usage through a compiler toolchain that leverages DMA-aware precision tuning and mathematical function memorization. The proposed solution extends the LLVM infrastructure, employing the TAFFO plugins for precision tuning, with the SETHET extension for DMA-aware precision tuning and LUTHET for automated, DMA-aware mathematical function memorization. We performed an experimental assessment on HERO, a heterogeneous platform employing RISC-V cores as a parallel accelerator. Our solution enables speedups ranging from 1.5× to 51.1× on AxBench benchmarks that employ trigonometrical functions and 4.23–48.4× on Polybench benchmarks over the baseline HERO platform.

Most Wearable Ultrasound (WUS) devices lack the computational power to process signals at the edge, instead relying on remote offload, which introduces latency, high power consumption, and privacy concerns. We present Maestro, a RISC-V SoC with unified Vector-Tensor Unit (VTU) and memory-coupled Fast Fourier Transform (FFT) accelerators targeting edge processing for wearable ultrasound devices, fabricated using low-cost TSMC 65nm CMOS technology. The VTU achieves peak 302GFLOPS/W and 19.8GFLOPS at FP16, while the multi-precision 16/32-bit floating-point FFT accelerator delivers peak 60.6GFLOPS/W and 3.6GFLOPS at FP16. We evaluate Maestro on a US-based gesture recognition task, achieving 1.62GFLOPS in signal processing at 26.68GFLOPS/W, and 19.52GFLOPS in Convolutional Neural Network (CNN) workloads at 298.03GFLOPS/W. Compared to a state-of-the-art SoC with a similar mission profile, Maestro achieves a$5\times $speedup while consuming only 12mW, with an energy consumption of 2.5mJ in a wearable US channel preprocessing and ML-based postprocessing pipeline.

QR decomposition is a numerical method used in many applications from the High-Performance Computing (HPC) domain to embedded systems. This broad spectrum of applications has drawn academic and commercial attention to developing many software libraries and domain-specific hardware solutions. In the Internet of Things (IoT) domain, multicore Parallel Ultra-Low-Power (PULP) architectures are emerging as energy-efficient alternatives, outperforming conventional single-core devices by coupling parallel processing with near-threshold computing. To the best of the authors' knowledge, our study introduces the first parallelized and optimized implementation of three distinct QR decomposition methods (Givens rotations, Gram-Schmidt process, and Householder transformation) on GAP-9, a commercial embodiment of the PULP architecture. Parallel execution on the 8-core cluster leads to a reduction in the total number of cycles by 241% for Givens rotations, 470% for Gram-Schmidt, and 567% for Householder, compared to the GAP9 1-core scenario. while each of them only consumes 0.013 mJ, 0.012 mJ, and 0.216 mJ, respectively. Compared to traditional single-core architectures based on ARM architectures, we achieve 8×, 24×, and 30× better performance and 36×, 35×, and 30× better energy efficiency, paving the way for broad adoption of complex linear algebra tasks in the IoT domain.

Energy and performance optimization of embedded hardware and software is of critical importance to achieve the overall system goals. In this work, we study the optimization of memory access through a combination of hardware (Direct Memory Access, DMA) and software (Precision Tuning) techniques, and we propose a compiler toolchain for managing both in the context of heterogeneous RISC-Vbased platforms. Our proposed toolchain, SeTHet, enables 3 - - 48 × speedup over the baseline system when employing both DMA and precision tuning, regardless of the availability of floating point units in hardware. SeTHet also achieves up to 16× speedup compared to DMA alone, thus proving that the combination of the two techniques provides a major improvement over either technique employed in isolation.

The demand for computation resources and energy efficiency of Convolutional Neural Networks (CNN) applications requires a new paradigm to overcome the “Memory Wall”. Analog In-Memory Computing (AIMC) is a promising paradigm since it performs matrix-vector multiplications, the critical kernel of many ML applications, in-place in the analog domain within memory arrays structured as crossbars of memory cells. However, several factors limit the full exploitation of this technology, including the physical fabrication of the crossbar devices, which constrain the memory capacity of a single array. Multi-AIMC architectures have been proposed to overcome this limitation, but they have been demonstrated only for tiny and custom CNNs or performing some layers off-chip. In this work, we present the full inference of an end-to-end ResNet-18 DNN on a 512-cluster heterogeneous architecture coupling a mix of AIMC cores and digital RISC-V cores, achieving up to 20.2 TOPS. Moreover, we analyze the mapping of the network on the available non-volatile cells, compare it with state-of-the-art models, and derive guidelines for next-generation many-core architectures based on AIMC devices.

Reduced-precision floating-point (FP) arithmetic is being widely adopted to reduce memory footprint and execution time on battery-powered Internet of Things (IoT) end-nodes. However, reduced precision computations must meet end-do-end precision constraints to be acceptable at the application level. This work introduces TransLib11https://github.com/ahmad-mirsalari/TransLib, an open-source kernel library based on transprecision computing principles, which provides knobs to exploit different FP data types (i.e., float, float16, and bfloat16), also considering the trade-off between homogeneous and mixed-precision solutions. We demonstrate the capabilities of the proposed library on PULP, a 32-bit microcontroller (MCU) coupled with a parallel, programmable accelerator. On average, TransLib kernels achieve an IPC of 0.94 and a speed-up of$1.64\times$using 16-bit vectorization. The parallel variants achieve a speed-up of$1.97\times, 3.91\times$, and$7.59\times$on 2, 4, and 8 cores, respectively. The memory footprint reduction is between 25% and 50%. Finally, we show that mixed-precision variants increase the accuracy by$30\times$at the cost of$2.09\times$execution time and$1.35\times$memory footprint compared to float16 vectorized.

IoT applications span a wide range in performance and memory footprint, under tight cost and power constraints. High-end applications rely on power-hungry Systems-on-Chip (SoCs) featuring powerful processors, large LPDDR/DDR3/4/5 memories, and supporting full-fledged Operating Systems (OS). On the contrary, low-end applications typically rely on Ultra-Low-Power μcontrollers with a “close to metal” software en-vironment and simple micro-kernel-based runtimes. Emerging applications and trends of IoT require the “best of both worlds”: cheap and low-power SoC systems with a well-known and agile software environment based on full-fledged OS (e.g., Linux), coupled with extreme energy efficiency and parallel digital signal processing capabilities. We present HULK-V: an open-source Heterogeneous Linux-capable RISC-V-based SoC coupling a 64-bit RISC-V processor with an 8-core Programmable Multi-Core Accelerator (PMCA), delivering up to 13.8 GOps, up to 157 GOps/W and accelerating the execution of complex DSP and ML tasks by up to 112× over the host processor. HULK-V leverages a lightweight, fully digital memory hierarchy based on HyperRAM IoT DRAM that exposes up to 512 MB of DRAM memory to the host CPU. Featuring HyperRAMs, HULK-V doubles the energy efficiency without significant performance loss compared to featuring power-hungry LPDDR memories, requiring expensive and large mixed-signal PHYs. HULK-V, implemented in Global Foundries 22nm FDX technology, is a fully digital ultra-low-cost SoC running a 64-bit Linux software stack with OpenMP host-to-PMCA offload within a power envelope of just 250 mW.

Computationally intensive algorithms such as Deep Neural Networks (DNNs) are becoming killer applications for edge devices. Porting heavily data-parallel algorithms on resource-constrained and battery-powered devices while retaining the flexibility granted by instruction processor-based architectures poses several challenges related to memory footprint, computational throughput, and energy efficiency. Low-bitwidth and mixed-precision arithmetic have been proven to be valid strategies for tackling these problems. We present Dustin, a fully programmable compute cluster integrating 16 RISC-V cores capable of 2- to 32-bit arithmetic and all possible mixed-precision combinations. In addition to a conventional Multiple-Instruction Multiple-Data (MIMD) processing paradigm, Dustin introduces a Vector Lockstep Execution Mode (VLEM) to minimize power consumption in highly data-parallel kernels. In VLEM, a single leader core fetches instructions and broadcasts them to the 15 follower cores. Clock gating Instruction Fetch (IF) stages and private caches of the follower cores leads to 38% power reduction. The cluster, implemented in 65 nm CMOS technology, achieves a peak performance of 58 GOPS and a peak efficiency of 1.15 TOPS/W.

Machine Learning (ML) functions are becoming ubiquitous in latency- and privacy-sensitive IoT applications, prompting a shift toward near-sensor processing at the extreme edge and the consequent increasing adoption of Parallel Ultra-low-power (PULP) IoT processors. These compute- and memory-constrained parallel architectures need to run efficiently a wide range of algorithms, including key Non-neural ML kernels that compete favorably with Deep Neural Networks in terms of accuracy under severe resource constraints. In this article, we focus on enabling efficient parallel execution of Non-neural ML algorithms on two RISCV-based PULP platforms, namely, GAP8, a commercial chip, and PULP-OPEN, a research platform running on an FPGA emulator. We optimized the parallel algorithms through a fine-grained analysis and intensive optimization to maximize the speedup, considering two alternative Floating-point (FP) emulation libraries on GAP8 and the native FPU support on PULP-OPEN. Experimental results show that a target-optimized emulation library can lead to an average 1.61× runtime improvement and 37% energy reduction compared to a standard emulation library, while the native FPU support reaches up to 32.09× and 99%, respectively. In terms of parallel speedup, our design improves the sequential execution by 7.04× on average on the targeted octa-core platforms leading to energy and latency decrease up to 87%. Last, we present a comparison with the ARM Cortex-M4 microcontroller, a widely adopted commercial solution for edge deployments, which is 12.87× slower than PULP-OPEN.

High performance and energy efficiency are critical requirements for Internet of Things (IoT) end-nodes. Exploiting tightly coupled clusters of programmable processors (CMPs) has recently emerged as a suitable solution to address this challenge. One of the main bottlenecks limiting the performance and energy efficiency of these systems is the instruction cache architecture due to its criticality in terms of timing (i.e., maximum operating frequency), bandwidth, and power. We propose a hierarchical instruction cache tailored to ultralow-power (ULP) tightly coupled processor clusters where a relatively large cache (L1.5) is shared by L1 private (PR) caches through a two-cycle latency interconnect. To address the performance loss caused by the L1 capacity misses, we introduce a next-line prefetcher with cache probe filtering (CPF) from L1 to L1.5. We optimize the core instruction fetch (IF) stage by removing the critical core-to-L1 combinational path. We present a detailed comparison of instruction cache architectures’ performance and energy efficiency for parallel ULP (PULP) clusters. Focusing on the implementation, our two-level instruction cache provides better scalability than existing shared caches, delivering up to 20% higher operating frequency. On average, the proposed two-level cache improves maximum performance by up to 17% compared to the state-of-the-art while delivering similar energy efficiency for most relevant applications.

Random forests (RFs) use a collection of decision trees (DTs) to perform the classification or regression. RFs are adopted in a wide variety of machine learning (ML) applications, and they are finding increasing use also in scenarios at the extreme edge of the Internet of Things (TinyML) where memory constraints are particularly tight. This article addresses the optimization of the computational and storage costs for running DTs on the microcontroller units (MCUs) typically deployed in TinyML scenarios. We introduce three alternative DT kernels optimized for memory- and compute-limited MCUs, providing insight into the key memory-latency tradeoffs on an open-source RISC-V platform. We identify key bottlenecks and demonstrate that SW optimizations enable up to significant memory footprint and latency decrease. Experimental results show that the optimized kernels achieve up to 4.5$\mu \text{s}$latency,$4.8\times $speedup, and 45% storage reduction against the widely-adopted naive DT design. We carry out a detailed performance and energy cost analysis of various optimized DT variants: the best approach requires just 8 instructions and 0.155 pJ per decision.

Recent applications in low-power (1-20 mW) near-sensor computing require the adoption of floating-point arithmetic to reconcile high precision results with a wide dynamic range. In this article, we propose a low-power multi-core computing cluster that leverages the fined-grained tunable principles of transprecision computing to provide support to near-sensor applications at a minimum power budget. Our solution – based on the open-source RISC-V architecture – combines parallelization and sub-word vectorization with a dedicated interconnect design capable of sharing floating-point units (FPUs) among the cores. On top of this architecture, we provide a full-fledged software stack support, including a parallel low-level runtime, a compilation toolchain, and a high-level programming model, with the aim to support the development of end-to-end applications. We performed an exhaustive exploration of the design space of the transprecision cluster on a cycle-accurate FPGA emulator, varying the number of cores and FPUs to maximize performance. Orthogonally, we performed a vertical exploration to identify the most efficient solutions in terms of non-functional requirements (operating frequency, power, and area). We conducted an experimental assessment on a set of benchmarks representative of the near-sensor processing domain, complementing the timing results with a post place-&-route analysis of the power consumption. A comparison with the state-of-the-art shows that our solution outperforms the competitors in energy efficiency, reaching a peak of 97 Gflop/s/W on single-precision scalars and 162 Gflop/s/W on half-precision vectors. Finally, a real-life use case demonstrates the effectiveness of our approach in fulfilling accuracy constraints.

Published in "IEEE Transactions on Emerging Topics in Computing, Volume: 9, Issue: 3, JulySeptember 2021" and orally presented at ARITH 2021.

The analysis of source code through machine learning techniques is an increasingly explored research topic aiming at increasing smartness in the software toolchain to exploit modern architectures in the best possible way. In the case of low-power, parallel embedded architectures, this means finding the configuration, for instance in terms of the number of cores, leading to minimum energy consumption. Depending on the kernel to be executed, the energy optimal scaling configuration is not trivial. While recent work has focused on general-purpose systems to learn and predict the best execution target in terms of the execution time of a snippet of code or kernel (e.g. offload OpenCL kernel on multicore CPU or GPU), in this work we focus on static compile-time features to assess if they can be successfully used to predict the minimum energy configuration on PULP, an ultra-low-power architecture featuring an on-chip cluster of RISC-V processors. Experiments show that using machine learning models on the source code to select the best energy scaling configuration automatically is viable and has the potential to be used in the context of automatic system configuration for energy minimisation.

The last few years have seen the emergence of IoT processors: ultra-low power systems-on-chips (SoCs) combining lightweight and flexible micro-controller units (MCUs), often based on open-ISA RISC-V cores, with application-specific accelerators to maximize performance and energy efficiency. Overall, this heterogeneity level requires complex hardware and a full-fledged software stack to orchestrate the execution and exploit platform features. For this reason, enabling agile design space exploration becomes a crucial asset for this new class of low-power SoCs. In this scenario, high-level simulators play an essential role in breaking the speed and design effort bottlenecks of cycle-accurate simulators and FPGA prototypes, respectively, while preserving functional and timing accuracy. We present GVSoC, a highly configurable and timing-accurate event-driven simulator that combines the efficiency of C++ models with the flexibility of Python configuration scripts. GVSoC is fully open-sourced, with the intent to drive future research in the area of highly parallel and heterogeneous RISC-V based IoT processors, leveraging three foundational features: Python-based modular configuration of the hardware description, easy calibration of platform parameters for accurate performance estimation, and high-speed simulation. Experimental results show that GVSoC enables practical functional and performance analysis and design exploration at the full-platform level (processors, memory, peripherals and IOs) with a speed-up of 2500× with respect to cycle-accurate simulation with errors typically below 10% for performance analysis.

The deployment of Deep Neural Networks (DNNs) on end-nodes at the extreme edge of the Internet-of-Things is a critical enabler to support pervasive Deep Learning-enhanced applications. Low-Cost MCU-based end-nodes have limited on-chip memory and often replace caches with scratchpads, to reduce area overheads and increase energy efficiency - requiring explicit DMA-based memory transfers between different levels of the memory hierarchy. Mapping modern DNNs on these systems requires aggressive topology-dependent tiling and double-buffering. In this work, we propose DORY (Deployment Oriented to memoRY) - an automatic tool to deploy DNNs on low cost MCUs with typically less than 1MB of on-chip SRAM memory. DORY abstracts tiling as a Constraint Programming (CP) problem: it maximizes L1 memory utilization under the topological constraints imposed by each DNN layer. Then, it generates ANSI C code to orchestrate off- and on-chip transfers and computation phases. Furthermore, to maximize speed, DORY augments the CP formulation with heuristics promoting performance-effective tile sizes. As a case study for DORY, we target GreenWaves Technologies GAP8, one of the most advanced parallel ultra-low power MCU-class devices on the market. On this device, DORY achieves up to 2.5× better MAC/cycle than the GreenWaves proprietary software solution and 18.1× better than the state-of-the-art result on an STM32-H743 MCU on single layers. Using our tool, GAP-8 can perform end-to-end inference of a 1.0-MobileNet-128 network consuming just 63 pJ/MAC on average @ 4.3 fps - 15.4× better than an STM32-H743. We release all our developments - the DORY framework, the optimized backend kernels, and the related heuristics - as open-source software.

The steeply growing performance demands for highly power- and energy-constrained processing systems such as end-nodes of the Internet-of-Things (IoT) have led to parallel near-threshold computing (NTC), joining the energy-efficiency benefits of low-voltage operation with the performance typical of parallel systems. Shared-L1-memory multiprocessor clusters are a promising architecture, delivering performance in the order of GOPS and over 100 GOPS/W of energy-efficiency. However, this level of computational efficiency can only be reached by maximizing the effective utilization of the processing elements (PEs) available in the clusters. Along with this effort, the optimization of PE-to-PE synchronization and communication is a critical factor for performance. In this article, we describe a light-weight hardware-accelerated synchronization and communication unit (SCU) for tightly-coupled clusters of processors. We detail the architecture, which enables fine-grain per-PE power management, and its integration into an eight-core cluster of RISC-V processors. To validate the effectiveness of the proposed solution, we implemented the eight-core cluster in advanced 22 nm FDX technology and evaluated performance and energy-efficiency with tunable microbenchmarks and a set of real-life applications and kernels. The proposed solution allows synchronization-free regions as small as 42 cycles, over 41× smaller than the baseline implementation based on fast test-and-set access to L1 memory when constraining the microbenchmarks to 10 percent synchronization overhead. When evaluated on the real-life DSP-applications, the proposed SCU improves performance by up to 92 and 23 percent on average and energy efficiency by up to 98 and 39 percent on average.

The growing demands of the worldwide IT infrastructure stress the need for reduced power consumption, which is addressed in so-called transprecision computing by improving energy efficiency at the expense of precision. For example, reducing the number of bits for some floating-point operations leads to higher efficiency, but also to a non-linear decrease of the computation accuracy. Depending on the application, small errors can be tolerated, thus allowing to fine-tune the precision of the computation. Finding the optimal precision for all variables in respect of an error bound is a complex task, which is tackled in the literature via heuristics. In this paper, we report on a first attempt to address the problem by combining a Mathematical Programming (MP) model and a Machine Learning (ML) model, following the Empirical Model Learning methodology. The ML model learns the relation between variables precision and the output error; this information is then embedded in the MP focused on minimizing the number of bits. An additional refinement phase is then added to improve the quality of the solution. The experimental results demonstrate an average speedup of 6.5% and a 3% increase in solution quality compared to the state-of-the-art. In addition, experiments on a hardware platform capable of mixed-precision arithmetic (PULPissimo) show the benefits of the proposed approach, with energy savings of around 40% compared to fixed-precision.

The deployment of Quantized Neural Networks (QNN) on advanced microcontrollers requires optimized software to exploit digital signal processing (DSP) extensions of modern instruction set architectures (ISA). As such, recent research proposed optimized libraries for QNNs (from 8-bit to 2-bit) such as CMSIS-NN and PULP-NN. This work presents an extension to the PULP-NN library targeting the acceleration of mixed-precision Deep Neural Networks, an emerging paradigm able to significantly shrink the memory footprint of deep neural networks with negligible accuracy loss. The library, composed of 27 kernels, one for each permutation of input feature maps, weights, and output feature maps precision (considering 8-bit, 4-bit and 2-bit), enables efficient inference of QNN on parallel ultra-low-power (PULP) clusters of RISC-V based processors, featuring the RV32IMCXpulpV2 ISA. The proposed solution, benchmarked on an 8-cores GAP-8 PULP cluster, reaches peak performance of 16 MACs/cycle on 8 cores, performing 21× to 25× faster than an STM32H7 (powered by an ARM Cortex M7 processor) with 15× to 21× better energy efficiency.

Strongly quantized fixed-point arithmetic is considered the key direction to enable the inference of CNNs on low-power, resource-constrained edge devices. However, the deployment of highly quantized Neural Networks at the extreme edge of IoT, on fully programmable MCUs, is currently limited by the lack of support, at the Instruction Set Architecture (ISA) level, for sub-byte fixed-point data types, making it necessary to add numerous instructions for packing and unpacking data when running low-bitwidth (i.e. 2- and 4-bit) QNN kernels, creating a bottleneck for performance and energy efficiency of QNN inference. In this work we present a set of extensions to the RISC-V ISA, aimed at boosting the energy efficiency of low-bitwidth QNNs on low-power microcontroller-class cores. The microarchitecture supporting the new extensions is built on top of a RISC-V core featuring instruction set extensions targeting energy-efficient digital signal processing. To evaluate the extensions, we integrated the core into a full microcontroller system, synthesized and placed&routed in 22nm FDX technology. QNN convolution kernels, implemented on the new core, run 5.3× and 8.9× faster when considering 4- and 2-bit data operands respectively, compared to the baseline processor only supporting 8-bit SIMD instructions. With a peak of 279 GMAC/s/W, the proposed solution achieves 9× better energy efficiency compared to the baseline and two orders of magnitudes better energy efficiency compared to state-of-the-art microcontrollers.

In recent years, Coarse Grain Reconfigurable Architecture (CGRA) accelerators have been increasingly deployed in Internet-of-Things (IoT) end nodes. A modern CGRA has to support and efficiently accelerate both integer and floating-point (FP) operations. In this paper, we propose an ultra-low-power tunable-precision CGRA architectural template, called TRANSprecision floating-point Programmable archItectuRE (TRANSPIRE), and its associated compilation flow supporting both integer and FP operations. TRANSPIRE employs transprecision computing and multiple Single Instruction Multiple Data (SIMD) to accelerate FP operations while boosting energy efficiency as well. Experimental results show that TRANSPIRE achieves a maximum of 10.06× performance gain and consumes 12.91× less energy w.r.t. a RISC-V based CPU with an enhanced ISA supporting SIMD-style vectorization and FP data-types, while executing applications for near-sensor computing and embedded machine learning, with an area overhead of 1.25× only.

In recent years approximate computing has been extensively explored as a paradigm to design hardware and software solutions that save energy by trading off on the quality of the computed results. In applications that involve numerical computations with wide dynamic range, precision tuning of floating-point (FP) variables is a key knob to leverage the energy/quality tradeoff of program results. This aspect assumes maximum relevance in the transprecision computing scenario, where accuracy of data is tuned at fine grain in application code. Performing precision tuning at fine grain requires a software development flow that streamlines the assessment of which variables have “precision slack” within an application. In this paper, we introduce FlexFloat, an open-source software library that has been expressly designed to aid the development of transprecision applications. FlexFloat provides a C/C++ interface for supporting multiple FP formats. Unlike alternative libraries, FlexFloat enables to control the bit-width of mantissa and exponent fields and provides advanced features for the collection of runtime statistics, reducing the FP emulation time compared to the state-of-the-art solutions. Its design allows to emulate the behavior of standard IEEE FP types and custom extensions for reduced-precision computation. This makes the library suitable for adoption in multiple contexts, from manual exploration to integration into automatic tools. Experimental findings demonstrate that our approach can be used to perform a complete precision analysis from which deriving multiple program versions depending on the energy/quality tradeoff. Furthermore, we show that the adoption of our methodology can lead to a significant reduction of energy consumption even on current commercial hardware (an embedded GPGPU).

RISC-V is an open-source instruction set architecture (ISA) with a modular design consisting of a mandatory base part plus optional extensions. The RISC-V 32IMFC ISA configuration has been widely adopted for the design of new-generation, low-power processors. Motivated by the important energy savings that smaller-than-32-bit FP types have enabled in several application domains and related compute platforms, some recent studies have published encouraging early results for their adoption in RISC-V processors. In this paper we introduce a set of ISA extensions for RISC-V 32IMFC, supporting scalar and SIMD operations (fitting the 32-bit register size) for 8-bit and two 16-bit FP types. The proposed extensions are enabled by exposing the new FP types to the standard C/C++ type system and an implementation for the RISC-V GCC compiler is presented. As a further, novel contribution, we extensively characterize the performance and energy savings achievable with the proposed extensions. On average, experimental results show that their adoption provide benefits in terms of performance (1.64× speedup for 16-bit and 2.18× for 8-bit types) and energy consumption (30% saving for 16-bit and 50% for 8-bit types). We also illustrate an approach based on automatic precision tuning to make effective use of the new FP types.

Guaranteed numerical precision of each elementary step in a complex computation has been the mainstay of traditional computing systems for many years. This era, fueled by Moore's law and the constant exponential improvement in computing efficiency, is at its twilight: from tiny nodes of the Internet-of-Things, to large HPC computing centers, sub-picoJoule/operation energy efficiency is essential for practical realizations. To overcome the power wall, a shift from traditional computing paradigms is now mandatory. In this paper we present the driving motivations, roadmap, and expected impact of the European project OPRECOMP. OPRECOMP aims to (i) develop the first complete transprecision computing framework, (ii) apply it to a wide range of hardware platforms, from the sub-milliWatt up to the MegaWatt range, and (iii) demonstrate impact in a wide range of computational domains, spanning IoT, Big Data Analytics, Deep Learning, and HPC simulations. By combining together into a seamless design transprecision advances in devices, circuits, software tools, and algorithms, we expect to achieve major energy efficiency improvements, even when there is no freedom to relax end-to-end application quality of results. Indeed, OPRECOMP aims at demolishing the ultra-conservative “precise” computing abstraction, replacing it with a more flexible and efficient one, namely transprecision computing.

In modern low-power embedded platforms, the execution of floating-point (FP) operations emerges as a major contributor to the energy consumption of compute-intensive applications with large dynamic range. Experimental evidence shows that 50% of the energy consumed by a core and its data memory is related to FP computations. The adoption of FP formats requiring a lower number of bits is an interesting opportunity to reduce energy consumption, since it allows to simplify the arithmetic circuitry and to reduce the memory bandwidth required to transfer data between memory and registers by enabling vectorization. From a theoretical point of view, the adoption of multiple FP types perfectly fits with the principle of transprecision computing, allowing fine-grained control of approximation while meeting specified constraints on the precision of final results. In this paper we propose an extended FP type system with complete hardware support to enable transprecision computing on low-power embedded processors, including two standard formats (binary32 and binary16) and two new formats (binary8 and binary16alt). First, we introduce a software library that enables exploration of FP types by tuning both precision and dynamic range of program variables. Then, we present a methodology to integrate our library with an external tool for precision tuning, and experimental results that highlight the clear benefits of introducing the new formats. Finally, we present the design of a transprecision FP unit capable of handling 8-bit and 16-bit operations in addition to standard 32-bit operations. Experimental results on FP-intensive benchmarks show that up to 90% of FP operations can be safely scaled down to 8-bit or 16-bit formats. Thanks to precision tuning and vectorization, execution time is decreased by 12% and memory accesses are reduced by 27% on average, leading to a reduction of energy consumption up to 30%.

Ultra-low power computing is a key enabler of deeply embedded platforms used in domains such as distributed sensing, internet of things, wearable computing. The rising computational demands and high dynamic of target algorithms often call for hardware support of floating-point (FP) arithmetic and high system energy efficiency. In light of transprecision computing, where accuracy of data is consciously changed during the execution of applications, custom FP types are being used to optimize a wide range of problems. We support two such custom types - one 16 bit and one 8 bit wide - together with IEEE binary16 as a set of "smallFloat" formats. We present an FP arithmetic unit capable of performing basic operations on smallFloat formats as well as conversions. To boost performance and energy efficiency, the smallFloat unit is extended with SIMD-style vectorization support to operate on a conventional word width of 32 bit. Finally, it is added into the execution stage of a low-power 32-bit RISC-V processor core and integrated as part of an SoC in a 65nm process. We show that the energy efficiency for processing smallFloat data in this amended system is 18% higher than the binary32 baseline, thus enabling hardware-supported power savings for applications making use of transprecision.

Ultralow-power embedded systems have recently started the move to multicore designs. Aggressive voltage scaling techniques have the potential to reduce the power consumption within the admitted envelope, but memory operations on standard six-transistor static RAM (6T-SRAM) cells become unreliable at low voltages. While standard cell memory (SCM) overcomes this limitation, it has much lower area density than SRAM, and thus it is too costly. On the other hand, several applications have inherent tolerance to computation errors, and executing such workloads with approximation has already proven a viable way to reduce energy consumption. In this paper, we propose a novel HW/SW approach to design energy-efficient ultralow-power systems which combine the key ideas of approximate computing and hybrid memory systems featuring both SCM and 6T-SRAM. We introduce a novel hardware support to split error-tolerant data so to host most significant bits in the SCM and least significant bits (LSBs) in the 6T-SRAM. This allows to power the memory system at a low voltage while ensuring correct operation by binding possible (flip-bit) errors to the LSBs only. In addition, by organizing 6T-SRAM banks into multiple and independent voltage domains we enable fine-grained, software-controlled voltage switching policies. At the software level, we propose language constructs to specify what regions of code and what variables are tolerant to approximation, plus compiler support to optimize data placement. Experimental results show that our proposal can reduce the energy consumption of the memory system by 47% on average, always complying with the result accuracy required by practical applications constraints.

In recent years, programmable many-core accelerators (PMCAs) have been introduced in embedded systems to satisfy stringent performance/Watt requirements. This has increased the urge for programming models capable of effectively leveraging hundreds to thousands of processors. Task-based parallelism has the potential to provide such capabilities, offering high-level abstractions to outline abundant and irregular parallelism in embedded applications. However, efficiently supporting this programming paradigm on embedded PMCAs is challenging, due to the large time and space overheads it introduces. In this paper we describe a lightweight OpenMP tasking runtime environment (RTE) design for a state-of-the-art embedded PMCA, the Kalray MPPA 256. We provide an exhaustive characterization of the costs of our RTE, considering both synthetic workload and real programs, and we compare to several other tasking RTEs. Experimental results confirm that our solution achieves near-ideal parallelization speedups for tasks as small as 5K cycles, and an average speedup of 12x for real benchmarks, which is 60% higher than what we observe with the original Kalray OpenMP implementation.

mW-scale parallel accelerators are a promising target for application domains such as the Internet of Thing (IoT), which require a strong compliance with a limited power budget combined with high performance capabilities. An important use case is given by smart sensing devices featuring increasingly sophisticated vision capabilities, at the cost of an increasing amount of near-sensor computation power. OpenVX is an emerging standard for the embedded vision, and provides a C-based application programming interface and a runtime environment. OpenVX is designed to maximize functional and performance portability across diverse hardware platforms. However, state-of-the-art implementations rely on memory-hungry data structures, which cannot be supported in constrained devices. In this paper we propose an alternative and novel approach to provide OpenVX support in mW-scale parallel accelerators. Our main contributions are: (i) an extension to the original OpenVX model to support static management of application graphs in the form of binary files; (ii) the definition of a companion runtime environment providing a lightweight support to execute binary graphs in a resource-constrained environment. Our approach achieves 68% memory footprint reduction and 3× execution speed-up compared to a baseline implementation. At the same time, data memory bandwidth is reduced by 10% and energy efficiency is improved by 2×.

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.

Nowadays Computer Vision application are ubiquitous, and their presence on embedded devices is more and more widespread. Heterogeneous embedded systems featuring a clustered manycore accelerator are a very promising target to execute embedded vision algorithms, but the code optimization for these platforms is a challenging task. Moreover, designers really need support tools that are both fast and accurate. In this work we introduce ADRENALINE, an environment for development and optimization of OpenVX applications targeting manycore accelerators. ADRENALINE consists of a custom OpenVX run-time and a virtual platform, and overall it is intended to provide support to enhance performance of embedded vision applications.

Multiprocessor systems-on-chip (MPSoC) are evolving into heterogeneous architectures based on one host processor plus many-core accelerators. While heterogeneous SoCs promise higher performance/watt, they are programmed at the cost of major code rewrites with low-level programming abstractions (e.g, OpenCL). We present a programming model based on OpenMP, with additional directives to program the accelerator from a single host program. As a test case, we evaluate an implementation of this programming model for the STMicroelectronics STHORM development board. We obtain near-ideal throughput for most benchmarks, very close performance to hand-optimized OpenCL codes at a significantly lower programming complexity, and up to 30× speedup versus host execution time.

Modern designs for embedded systems are increasingly embracing cluster-based architectures, where small sets of cores communicate through tightly-coupled shared memory banks and high-performance interconnections. At the same time, the complexity of modern applications requires new programming abstractions to exploit dynamic and/or irregular parallelism on such platforms. Supporting dynamic parallelism in systems which i) are resource-constrained and ii) run applications with small units of work calls for a runtime environment which has minimal overhead for the scheduling of parallel tasks. In this work, we study the major sources of overhead in the implementation of OpenMP dynamic loops, sections and tasks, and propose a hardware implementation of a generic Scheduling Engine (HWSE) which fits the semantics of the three constructs. The HWSE is designed as a tightly-coupled block to the PEs within a multi-core cluster, communicating through a shared-memory interface. This allows very fast programming and synchronization with the controlling PEs, fundamental to achieving fast dynamic scheduling, and ultimately to enable fine-grained parallelism. We prove the effectiveness of our solutions with real applications and synthetic benchmarks, using a cycle-accurate virtual platform.

Cluster-based architectures are increasingly being adopted to design embedded many-cores. These platforms can deliver very high peak performance within a contained power envelope, provided that programmers can make effective use the available parallel cores. This is becoming an extremely difficult task, as embedded applications are growing in complexity and exhibit irregular and dynamic parallelism. The OpenMP tasking extensions represent a powerful abstraction to capture this form of parallelism. However, efficiently supporting it on cluster-based embedded SoCs is not easy, because the fine-grained parallel workload present in embedded applications can not tolerate high memory and run-time overheads. In this paper we present our design of the runtime support layer to OpenMP tasking for an embedded shared memory cluster, identifying key aspects to achieving performance and discussing important architectural support to removing major bottlenecks.