Francesco Conti

Alessio Burrello, Alberto Dequino, Daniele Jahier Pagliari et al.

Temporal Convolutional Networks (TCNs) are emerging lightweight Deep Learning models for Time Series analysis. We introduce an automated exploration approach and a library of optimized kernels to map TCNs on Parallel Ultra-Low Power (PULP) microcontrollers. Our approach minimizes latency and energy by exploiting a layer tiling optimizer to jointly find the tiling dimensions and select among alternative implementations of the causal and dilated 1D-convolution operations at the core of TCNs. We benchmark our approach on a commercial PULP device, achieving up to 103X lower latency and 20.3X lower energy than the Cube-AI toolkit executed on the STM32L4 and from 2.9X to 26.6X lower energy compared to commercial closed-source and academic open-source approaches on the same hardware target.

Cristina Silvano, Daniele Ielmini, Fabrizio Ferrandi et al.

Recent trends in deep learning (DL) have made hardware accelerators essential for various high-performance computing (HPC) applications, including image classification, computer vision, and speech recognition. This survey summarizes and classifies the most recent developments in DL accelerators, focusing on their role in meeting the performance demands of HPC applications. We explore cutting-edge approaches to DL acceleration, covering not only GPU- and TPU-based platforms but also specialized hardware such as FPGA- and ASIC-based accelerators, Neural Processing Units, open hardware RISC-V-based accelerators, and co-processors. This survey also describes accelerators leveraging emerging memory technologies and computing paradigms, including 3D-stacked Processor-In-Memory, non-volatile memories like Resistive RAM and Phase Change Memories used for in-memory computing, as well as Neuromorphic Processing Units, and Multi-Chip Module-based accelerators. Furthermore, we provide insights into emerging quantum-based accelerators and photonics. Finally, this survey categorizes the most influential architectures and technologies from recent years, offering readers a comprehensive perspective on the rapidly evolving field of deep learning acceleration.

Yvan Tortorella, Luca Bertaccini, Luca Benini et al.

The increasing interest in TinyML, i.e., near-sensor machine learning on power budgets of a few tens of mW, is currently pushing toward enabling TinyML-class training as opposed to inference only. Current training algorithms, based on various forms of error and gradient backpropagation, rely on floating-point matrix operations to meet the precision and dynamic range requirements. So far, the energy and power cost of these operations has been considered too high for TinyML scenarios. This paper addresses the open challenge of near-sensor training on a few mW power budget and presents RedMulE - Reduced-Precision Matrix Multiplication Engine, a low-power specialized accelerator conceived for multi-precision floating-point General Matrix-Matrix Operations (GEMM-Ops) acceleration, supporting FP16, as well as hybrid FP8 formats, with {sign, exponent, mantissa}=({1,4,3}, {1,5,2}). We integrate RedMule into a Parallel Ultra-Low-Power (PULP) cluster containing eight energy-efficient RISC-V cores sharing a tightly-coupled data memory and implement the resulting system in a 22 nm technology. At its best efficiency point (@ 470 MHz, 0.65 V), the RedMulE-augmented PULP cluster achieves 755 GFLOPS/W and 920 GFLOPS/W during regular General Matrix-Matrix Multiplication (GEMM), and up to 1.19 TFLOPS/W and 1.67 TFLOPS/W when executing GEMM-Ops, respectively, for FP16 and FP8 input/output tensors. In its best performance point (@ 613 MHz, 0.8 V), RedMulE achieves up to 58.5 GFLOPS and 117 GFLOPS for FP16 and FP8, respectively, with 99.4% utilization of the array of Computing Elements and consuming less than 60 mW on average, thus enabling on-device training of deep learning models in TinyML application scenarios while retaining the flexibility to tackle other classes of common linear algebra problems efficiently.

Matteo Risso, Alessio Burrello, Daniele Jahier Pagliari et al.

Temporal Convolutional Networks (TCNs) are promising Deep Learning models for time-series processing tasks. One key feature of TCNs is time-dilated convolution, whose optimization requires extensive experimentation. We propose an automatic dilation optimizer, which tackles the problem as a weight pruning on the time-axis, and learns dilation factors together with weights, in a single training. Our method reduces the model size and inference latency on a real SoC hardware target by up to 7.4x and 3x, respectively with no accuracy drop compared to a network without dilation. It also yields a rich set of Pareto-optimal TCNs starting from a single model, outperforming hand-designed solutions in both size and accuracy.

Georg Rutishauser, Francesco Conti, Luca Benini

Mixed-precision quantization, where a deep neural network's layers are quantized to different precisions, offers the opportunity to optimize the trade-offs between model size, latency, and statistical accuracy beyond what can be achieved with homogeneous-bit-width quantization. To navigate the intractable search space of mixed-precision configurations for a given network, this paper proposes a hybrid search methodology. It consists of a hardware-agnostic differentiable search algorithm followed by a hardware-aware heuristic optimization to find mixed-precision configurations latency-optimized for a specific hardware target. We evaluate our algorithm on MobileNetV1 and MobileNetV2 and deploy the resulting networks on a family of multi-core RISC-V microcontroller platforms with different hardware characteristics. We achieve up to 28.6% reduction of end-to-end latency compared to an 8-bit model at a negligible accuracy drop from a full-precision baseline on the 1000-class ImageNet dataset. We demonstrate speedups relative to an 8-bit baseline, even on systems with no hardware support for sub-byte arithmetic at negligible accuracy drop. Furthermore, we show the superiority of our approach with respect to differentiable search targeting reduced binary operation counts as a proxy for latency.

Moritz Scherer, Luka Macan, Victor Jung et al.

With the rise of Embodied Foundation Models (EFMs), most notably Small Language Models (SLMs), adapting Transformers for edge applications has become a very active field of research. However, achieving end-to-end deployment of SLMs on microcontroller (MCU)-class chips without high-bandwidth off-chip main memory access is still an open challenge. In this paper, we demonstrate high-efficiency end-to-end SLM deployment on a multicore RISC-V (RV32) MCU augmented with ML instruction extensions and a hardware neural processing unit (NPU). To automate the exploration of the constrained, multi-dimensional memory vs. computation tradeoffs involved in aggressive SLM deployment on heterogeneous (multicore+NPU) resources, we introduce Deeploy, a novel Deep Neural Network (DNN) compiler, which generates highly-optimized C code requiring minimal runtime support. We demonstrate that Deeploy generates end-to-end code for executing SLMs, fully exploiting the RV32 cores' instruction extensions and the NPU: We achieve leading-edge energy and throughput of \SI{490}{\micro\joule \per Token}, at \SI{340}{Token \per \second} for an SLM trained on the TinyStories dataset, running for the first time on an MCU-class device without external memory.

Serena Curzel, Fabrizio Ferrandi, Leandro Fiorin et al.

Given their increasing size and complexity, the need for efficient execution of deep neural networks has become increasingly pressing in the design of heterogeneous High-Performance Computing (HPC) and edge platforms, leading to a wide variety of proposals for specialized deep learning architectures and hardware accelerators. The design of such architectures and accelerators requires a multidisciplinary approach combining expertise from several areas, from machine learning to computer architecture, low-level hardware design, and approximate computing. Several methodologies and tools have been proposed to improve the process of designing accelerators for deep learning, aimed at maximizing parallelism and minimizing data movement to achieve high performance and energy efficiency. This paper critically reviews influential tools and design methodologies for Deep Learning accelerators, offering a wide perspective in this rapidly evolving field. This work complements surveys on architectures and accelerators by covering hardware-software co-design, automated synthesis, domain-specific compilers, design space exploration, modeling, and simulation, providing insights into technical challenges and open research directions.

Philip Wiese, Gamze İslamoğlu, Moritz Scherer et al.

One of the challenges for Tiny Machine Learning (tinyML) is keeping up with the evolution of Machine Learning models from Convolutional Neural Networks to Transformers. We address this by leveraging a heterogeneous architectural template coupling RISC-V processors with hardwired accelerators supported by an automated deployment flow. We demonstrate Attention-based models in a tinyML power envelope with an octa-core cluster coupled with an accelerator for quantized Attention. Our deployment flow enables end-to-end 8-bit Transformer inference, achieving leading-edge energy efficiency and throughput of 2960 GOp/J and 154 GOp/s (0.65 V, 22 nm FD-SOI technology).

Luca Bompani, Manuele Rusci, Daniele Palossi et al.

This paper introduces Multi-Resolution Rescored Byte-Track (MR2-ByteTrack), a novel video object detection framework for ultra-low-power embedded processors. This method reduces the average compute load of an off-the-shelf Deep Neural Network (DNN) based object detector by up to 2.25$\times$ by alternating the processing of high-resolution images (320$\times$320 pixels) with multiple down-sized frames (192$\times$192 pixels). To tackle the accuracy degradation due to the reduced image input size, MR2-ByteTrack correlates the output detections over time using the ByteTrack tracker and corrects potential misclassification using a novel probabilistic Rescore algorithm. By interleaving two down-sized images for every high-resolution one as the input of different state-of-the-art DNN object detectors with our MR2-ByteTrack, we demonstrate an average accuracy increase of 2.16% and a latency reduction of 43% on the GAP9 microcontroller compared to a baseline frame-by-frame inference scheme using exclusively full-resolution images. Code available at: https://github.com/Bomps4/Multi_Resolution_Rescored_ByteTrack

Daniele Palossi, Antonio Loquercio, Francesco Conti et al.

Fully-autonomous miniaturized robots (e.g., drones), with artificial intelligence (AI) based visual navigation capabilities are extremely challenging drivers of Internet-of-Things edge intelligence capabilities. Visual navigation based on AI approaches, such as deep neural networks (DNNs) are becoming pervasive for standard-size drones, but are considered out of reach for nanodrones with size of a few cm${}^\mathrm{2}$. In this work, we present the first (to the best of our knowledge) demonstration of a navigation engine for autonomous nano-drones capable of closed-loop end-to-end DNN-based visual navigation. To achieve this goal we developed a complete methodology for parallel execution of complex DNNs directly on-bard of resource-constrained milliwatt-scale nodes. Our system is based on GAP8, a novel parallel ultra-low-power computing platform, and a 27 g commercial, open-source CrazyFlie 2.0 nano-quadrotor. As part of our general methodology we discuss the software mapping techniques that enable the state-of-the-art deep convolutional neural network presented in [1] to be fully executed on-board within a strict 6 fps real-time constraint with no compromise in terms of flight results, while all processing is done with only 64 mW on average. Our navigation engine is flexible and can be used to span a wide performance range: at its peak performance corner it achieves 18 fps while still consuming on average just 3.5% of the power envelope of the deployed nano-aircraft.

Victor J. B. Jung, Alessio Burrello, Moritz Scherer et al.

Transformer networks are rapidly becoming SotA in many fields, such as NLP and CV. Similarly to CNN, there is a strong push for deploying Transformer models at the extreme edge, ultimately fitting the tiny power budget and memory footprint of MCUs. However, the early approaches in this direction are mostly ad-hoc, platform, and model-specific. This work aims to enable and optimize the flexible, multi-platform deployment of encoder Tiny Transformers on commercial MCUs. We propose a complete framework to perform end-to-end deployment of Transformer models onto single and multi-core MCUs. Our framework provides an optimized library of kernels to maximize data reuse and avoid unnecessary data marshaling operations into the crucial attention block. A novel MHSA inference schedule, named Fused-Weight Self-Attention, is introduced, fusing the linear projection weights offline to further reduce the number of operations and parameters. Furthermore, to mitigate the memory peak reached by the computation of the attention map, we present a Depth-First Tiling scheme for MHSA. We evaluate our framework on three different MCU classes exploiting ARM and RISC-V ISA, namely the STM32H7, the STM32L4, and GAP9 (RV32IMC-XpulpV2). We reach an average of 4.79x and 2.0x lower latency compared to SotA libraries CMSIS-NN (ARM) and PULP-NN (RISC-V), respectively. Moreover, we show that our MHSA depth-first tiling scheme reduces the memory peak by up to 6.19x, while the fused-weight attention can reduce the runtime by 1.53x, and number of parameters by 25%. We report significant improvements across several Tiny Transformers: for instance, when executing a transformer block for the task of radar-based hand-gesture recognition on GAP9, we achieve a latency of 0.14ms and energy consumption of 4.92 micro-joules, 2.32x lower than the SotA PULP-NN library on the same platform.

Alberto Dequino, Alessio Carpegna, Davide Nadalini et al.

Rehearsal-based Continual Learning (CL) has been intensely investigated in Deep Neural Networks (DNNs). However, its application in Spiking Neural Networks (SNNs) has not been explored in depth. In this paper we introduce the first memory-efficient implementation of Latent Replay (LR)-based CL for SNNs, designed to seamlessly integrate with resource-constrained devices. LRs combine new samples with latent representations of previously learned data, to mitigate forgetting. Experiments on the Heidelberg SHD dataset with Sample and Class-Incremental tasks reach a Top-1 accuracy of 92.5% and 92%, respectively, without forgetting the previously learned information. Furthermore, we minimize the LRs' requirements by applying a time-domain compression, reducing by two orders of magnitude their memory requirement, with respect to a naive rehearsal setup, with a maximum accuracy drop of 4%. On a Multi-Class-Incremental task, our SNN learns 10 new classes from an initial set of 10, reaching a Top-1 accuracy of 78.4% on the full test set.

Davide Nadalini, Manuele Rusci, Elia Cereda et al.

Monocular depth estimation (MDE) plays a crucial role in enabling spatially-aware applications in Ultra-low-power (ULP) Internet-of-Things (IoT) platforms. However, the limited number of parameters of Deep Neural Networks for the MDE task, designed for IoT nodes, results in severe accuracy drops when the sensor data observed in the field shifts significantly from the training dataset. To address this domain shift problem, we present a multi-modal On-Device Learning (ODL) technique, deployed on an IoT device integrating a Greenwaves GAP9 MicroController Unit (MCU), a 80 mW monocular camera and a 8 x 8 pixel depth sensor, consuming $\approx$300mW. In its normal operation, this setup feeds a tiny 107 k-parameter $μ$PyD-Net model with monocular images for inference. The depth sensor, usually deactivated to minimize energy consumption, is only activated alongside the camera to collect pseudo-labels when the system is placed in a new environment. Then, the fine-tuning task is performed entirely on the MCU, using the new data. To optimize our backpropagation-based on-device training, we introduce a novel memory-driven sparse update scheme, which minimizes the fine-tuning memory to 1.2 MB, 2.2x less than a full update, while preserving accuracy (i.e., only 2% and 1.5% drops on the KITTI and NYUv2 datasets). Our in-field tests demonstrate, for the first time, that ODL for MDE can be performed in 17.8 minutes on the IoT node, reducing the root mean squared error from 4.9 to 0.6m with only 3 k self-labeled samples, collected in a real-life deployment scenario.

Run Wang, Gamze Islamoglu, Andrea Belano et al.

While Transformers are dominated by Floating-Point (FP) Matrix-Multiplications, their aggressive acceleration through dedicated hardware or many-core programmable systems has shifted the performance bottleneck to non-linear functions like Softmax. Accelerating Softmax is challenging due to its non-pointwise, non-linear nature, with exponentiation as the most demanding step. To address this, we design a custom arithmetic block for Bfloat16 exponentiation leveraging a novel approximation algorithm based on Schraudolph's method, and we integrate it into the Floating-Point Unit (FPU) of the RISC-V cores of a compute cluster, through custom Instruction Set Architecture (ISA) extensions, with a negligible area overhead of 1\%. By optimizing the software kernels to leverage the extension, we execute Softmax with 162.7$\times$ less latency and 74.3$\times$ less energy compared to the baseline cluster, achieving an 8.2$\times$ performance improvement and 4.1$\times$ higher energy efficiency for the FlashAttention-2 kernel in GPT-2 configuration. Moreover, the proposed approach enables a multi-cluster system to efficiently execute end-to-end inference of pre-trained Transformer models, such as GPT-2, GPT-3 and ViT, achieving up to 5.8$\times$ and 3.6$\times$ reduction in latency and energy consumption, respectively, without requiring re-training and with negligible accuracy loss.

Davide Nadalini, Manuele Rusci, Luca Benini et al.

Enabling On-Device Learning (ODL) for Ultra-Low-Power Micro-Controller Units (MCUs) is a key step for post-deployment adaptation and fine-tuning of Deep Neural Network (DNN) models in future TinyML applications. This paper tackles this challenge by introducing a novel reduced precision optimization technique for ODL primitives on MCU-class devices, leveraging the State-of-Art advancements in RISC-V RV32 architectures with support for vectorized 16-bit floating-point (FP16) Single-Instruction Multiple-Data (SIMD) operations. Our approach for the Forward and Backward steps of the Back-Propagation training algorithm is composed of specialized shape transform operators and Matrix Multiplication (MM) kernels, accelerated with parallelization and loop unrolling. When evaluated on a single training step of a 2D Convolution layer, the SIMD-optimized FP16 primitives result up to 1.72$\times$ faster than the FP32 baseline on a RISC-V-based 8+1-core MCU. An average computing efficiency of 3.11 Multiply and Accumulate operations per clock cycle (MAC/clk) and 0.81 MAC/clk is measured for the end-to-end training tasks of a ResNet8 and a DS-CNN for Image Classification and Keyword Spotting, respectively -- requiring 17.1 ms and 6.4 ms on the target platform to compute a training step on a single sample. Overall, our approach results more than two orders of magnitude faster than existing ODL software frameworks for single-core MCUs and outperforms by 1.6 $\times$ previous FP32 parallel implementations on a Continual Learning setup.

Francesco Conti, Gianna Paulin, Angelo Garofalo et al.

Emerging Artificial Intelligence-enabled Internet-of-Things (AI-IoT) System-on-a-Chip (SoC) for augmented reality, personalized healthcare, and nano-robotics need to run many diverse tasks within a power envelope of a few tens of mW over a wide range of operating conditions: compute-intensive but strongly quantized Deep Neural Network (DNN) inference, as well as signal processing and control requiring high-precision floating-point. We present Marsellus, an all-digital heterogeneous SoC for AI-IoT end-nodes fabricated in GlobalFoundries 22nm FDX that combines 1) a general-purpose cluster of 16 RISC-V Digital Signal Processing (DSP) cores attuned for the execution of a diverse range of workloads exploiting 4-bit and 2-bit arithmetic extensions (XpulpNN), combined with fused MAC&LOAD operations and floating-point support; 2) a 2-8bit Reconfigurable Binary Engine (RBE) to accelerate 3x3 and 1x1 (pointwise) convolutions in DNNs; 3) a set of On-Chip Monitoring (OCM) blocks connected to an Adaptive Body Biasing (ABB) generator and a hardware control loop, enabling on-the-fly adaptation of transistor threshold voltages. Marsellus achieves up to 180 Gop/s or 3.32 Top/s/W on 2-bit precision arithmetic in software, and up to 637 Gop/s or 12.4 Top/s/W on hardware-accelerated DNN layers.

Gianna Paulin, Francesco Conti, Lukas Cavigelli et al.

Recurrent neural networks such as Long Short-Term Memories (LSTMs) learn temporal dependencies by keeping an internal state, making them ideal for time-series problems such as speech recognition. However, the output-to-input feedback creates distinctive memory bandwidth and scalability challenges in designing accelerators for RNNs. We present Muntaniala, an RNN accelerator architecture for LSTM inference with a silicon-measured energy-efficiency of 3.25$TOP/s/W$ and performance of 30.53$GOP/s$ in UMC 65 $nm$ technology. The scalable design of Muntaniala allows running large RNN models by combining multiple tiles in a systolic array. We keep all parameters stationary on every die in the array, drastically reducing the I/O communication to only loading new features and sharing partial results with other dies. For quantifying the overall system power, including I/O power, we built Vau da Muntanialas, to the best of our knowledge, the first demonstration of a systolic multi-chip-on-PCB array of RNN accelerator. Our multi-die prototype performs LSTM inference with 192 hidden states in 330$μs$ with a total system power of 9.0$mW$ at 10$MHz$ consuming 2.95$μJ$. Targeting the 8/16-bit quantization implemented in Muntaniala, we show a phoneme error rate (PER) drop of approximately 3% with respect to floating-point (FP) on a 3L-384NH-123NI LSTM network on the TIMIT dataset.

Angelo Garofalo, Gianmarco Ottavi, Francesco Conti et al.

Deployment of modern TinyML tasks on small battery-constrained IoT devices requires high computational energy efficiency. Analog In-Memory Computing (IMC) using non-volatile memory (NVM) promises major efficiency improvements in deep neural network (DNN) inference and serves as on-chip memory storage for DNN weights. However, IMC's functional flexibility limitations and their impact on performance, energy, and area efficiency are not yet fully understood at the system level. To target practical end-to-end IoT applications, IMC arrays must be enclosed in heterogeneous programmable systems, introducing new system-level challenges which we aim at addressing in this work. We present a heterogeneous tightly-coupled clustered architecture integrating 8 RISC-V cores, an in-memory computing accelerator (IMA), and digital accelerators. We benchmark the system on a highly heterogeneous workload such as the Bottleneck layer from a MobileNetV2, showing 11.5x performance and 9.5x energy efficiency improvements, compared to highly optimized parallel execution on the cores. Furthermore, we explore the requirements for end-to-end inference of a full mobile-grade DNN (MobileNetV2) in terms of IMC array resources, by scaling up our heterogeneous architecture to a multi-array accelerator. Our results show that our solution, on the end-to-end inference of the MobileNetV2, is one order of magnitude better in terms of execution latency than existing programmable architectures and two orders of magnitude better than state-of-the-art heterogeneous solutions integrating in-memory computing analog cores.

Leonardo Ravaglia, Manuele Rusci, Davide Nadalini et al.

In the last few years, research and development on Deep Learning models and techniques for ultra-low-power devices in a word, TinyML has mainly focused on a train-then-deploy assumption, with static models that cannot be adapted to newly collected data without cloud-based data collection and fine-tuning. Latent Replay-based Continual Learning (CL) techniques[1] enable online, serverless adaptation in principle, but so farthey have still been too computation and memory-hungry for ultra-low-power TinyML devices, which are typically based on microcontrollers. In this work, we introduce a HW/SW platform for end-to-end CL based on a 10-core FP32-enabled parallel ultra-low-power (PULP) processor. We rethink the baseline Latent Replay CL algorithm, leveraging quantization of the frozen stage of the model and Latent Replays (LRs) to reduce their memory cost with minimal impact on accuracy. In particular, 8-bit compression of the LR memory proves to be almost lossless (-0.26% with 3000LR) compared to the full-precision baseline implementation, but requires 4x less memory, while 7-bit can also be used with an additional minimal accuracy degradation (up to 5%). We also introduce optimized primitives for forward and backward propagation on the PULP processor. Our results show that by combining these techniques, continual learning can be achieved in practice using less than 64MB of memory an amount compatible with embedding in TinyML devices. On an advanced 22nm prototype of our platform, called VEGA, the proposed solution performs onaverage 65x faster than a low-power STM32 L4 microcontroller, being 37x more energy efficient enough for a lifetime of 535h when learning a new mini-batch of data once every minute.

Davide Rossi, Francesco Conti, Manuel Eggimann et al.

The Internet-of-Things requires end-nodes with ultra-low-power always-on capability for a long battery lifetime, as well as high performance, energy efficiency, and extreme flexibility to deal with complex and fast-evolving near-sensor analytics algorithms (NSAAs). We present Vega, an IoT end-node SoC capable of scaling from a 1.7 $\mathrmμ$W fully retentive cognitive sleep mode up to 32.2 GOPS (@ 49.4 mW) peak performance on NSAAs, including mobile DNN inference, exploiting 1.6 MB of state-retentive SRAM, and 4 MB of non-volatile MRAM. To meet the performance and flexibility requirements of NSAAs, the SoC features 10 RISC-V cores: one core for SoC and IO management and a 9-cores cluster supporting multi-precision SIMD integer and floating-point computation. Vega achieves SoA-leading efficiency of 615 GOPS/W on 8-bit INT computation (boosted to 1.3TOPS/W for 8-bit DNN inference with hardware acceleration). On floating-point (FP) compuation, it achieves SoA-leading efficiency of 79 and 129 GFLOPS/W on 32- and 16-bit FP, respectively. Two programmable machine-learning (ML) accelerators boost energy efficiency in cognitive sleep and active states, respectively.

Daniele Palossi, Nicky Zimmerman, Alessio Burrello et al.

Artificial intelligence-powered pocket-sized air robots have the potential to revolutionize the Internet-of-Things ecosystem, acting as autonomous, unobtrusive, and ubiquitous smart sensors. With a few cm$^{2}$ form-factor, nano-sized unmanned aerial vehicles (UAVs) are the natural befit for indoor human-drone interaction missions, as the pose estimation task we address in this work. However, this scenario is challenged by the nano-UAVs' limited payload and computational power that severely relegates the onboard brain to the sub-100 mW microcontroller unit-class. Our work stands at the intersection of the novel parallel ultra-low-power (PULP) architectural paradigm and our general development methodology for deep neural network (DNN) visual pipelines, i.e., covering from perception to control. Addressing the DNN model design, from training and dataset augmentation to 8-bit quantization and deployment, we demonstrate how a PULP-based processor, aboard a nano-UAV, is sufficient for the real-time execution (up to 135 frame/s) of our novel DNN, called PULP-Frontnet. We showcase how, scaling our model's memory and computational requirement, we can significantly improve the onboard inference (top energy efficiency of 0.43 mJ/frame) with no compromise in the quality-of-result vs. a resource-unconstrained baseline (i.e., full-precision DNN). Field experiments demonstrate a closed-loop top-notch autonomous navigation capability, with a heavily resource-constrained 27-gram Crazyflie 2.1 nano-quadrotor. Compared against the control performance achieved using an ideal sensing setup, onboard relative pose inference yields excellent drone behavior in terms of median absolute errors, such as positional (onboard: 41 cm, ideal: 26 cm) and angular (onboard: 3.7$^{\circ}$, ideal: 4.1$^{\circ}$).

Alfio Di Mauro, Francesco Conti, Pasquale Davide Schiavone et al.

Binary Neural Networks (BNNs) have been shown to be robust to random bit-level noise, making aggressive voltage scaling attractive as a power-saving technique for both logic and SRAMs. In this work, we introduce the first fully programmable IoT end-node system-on-chip (SoC) capable of executing software-defined, hardware-accelerated BNNs at ultra-low voltage. Our SoC exploits a hybrid memory scheme where error-vulnerable SRAMs are complemented by reliable standard-cell memories to safely store critical data under aggressive voltage scaling. On a prototype in 22nm FDX technology, we demonstrate that both the logic and SRAM voltage can be dropped to 0.5Vwithout any accuracy penalty on a BNN trained for the CIFAR-10 dataset, improving energy efficiency by 2.2X w.r.t. nominal conditions. Furthermore, we show that the supply voltage can be dropped to 0.42V (50% of nominal) while keeping more than99% of the nominal accuracy (with a bit error rate ~1/1000). In this operating point, our prototype performs 4Gop/s (15.4Inference/s on the CIFAR-10 dataset) by computing up to 13binary ops per pJ, achieving 22.8 Inference/s/mW while keeping within a peak power envelope of 674uW - low enough to enable always-on operation in ultra-low power smart cameras, long-lifetime environmental sensors, and insect-sized pico-drones.

Francesco Conti

This technical report aims at defining a formal framework for Deep Neural Network (DNN) layer-wise quantization, focusing in particular on the problems related to the final deployment. It also acts as a documentation for the NEMO (NEural Minimization for pytOrch) framework. It describes the four DNN representations used in NEMO (FullPrecision, FakeQuantized, QuantizedDeployable and IntegerDeployable), focusing in particular on a formal definition of the latter two. An important feature of this model, and in particular the IntegerDeployable representation, is that it enables DNN inference using purely integers - without resorting to real-valued numbers in any part of the computation and without relying on an explicit fixed-point numerical representation.

Arthur Stoutchinin, Francesco Conti, Luca Benini

Embedded inference engines for convolutional networks must be parsimonious in memory bandwidth and buffer sizing to meet power and cost constraints. We present an analytical memory bandwidth model for loop-nest optimization targeting architectures with application managed buffers. We applied this model to optimize the CNN convolution loop-nest. We show that our model is more accurate than previously published models. Using this model we can identify non-trivial dataflow schedules that result in lowest communication bandwidth given tight local buffering constraints. We show that optimal dataflow schedules are implementable in practice and that our model is accurate with respect to a real implementation; moreover, we introduce an accelerator architecture, named Hardware Convolution Block (HWC), which implements the optimal schedules, and we show it achieves up to 14x memory bandwidth reduction compared to a previously published accelerator with a similar memory interface, but implementing a non-optimal schedule.

Francesco Conti, Pasquale Davide Schiavone, Luca Benini

Binary Neural Networks (BNNs) are promising to deliver accuracy comparable to conventional deep neural networks at a fraction of the cost in terms of memory and energy. In this paper, we introduce the XNOR Neural Engine (XNE), a fully digital configurable hardware accelerator IP for BNNs, integrated within a microcontroller unit (MCU) equipped with an autonomous I/O subsystem and hybrid SRAM / standard cell memory. The XNE is able to fully compute convolutional and dense layers in autonomy or in cooperation with the core in the MCU to realize more complex behaviors. We show post-synthesis results in 65nm and 22nm technology for the XNE IP and post-layout results in 22nm for the full MCU indicating that this system can drop the energy cost per binary operation to 21.6fJ per operation at 0.4V, and at the same time is flexible and performant enough to execute state-of-the-art BNN topologies such as ResNet-34 in less than 2.2mJ per frame at 8.9 fps.

Francesco Conti, Lukas Cavigelli, Gianna Paulin et al.

Recurrent neural networks (RNNs) are state-of-the-art in voice awareness/understanding and speech recognition. On-device computation of RNNs on low-power mobile and wearable devices would be key to applications such as zero-latency voice-based human-machine interfaces. Here we present Chipmunk, a small (<1 mm${}^2$) hardware accelerator for Long-Short Term Memory RNNs in UMC 65 nm technology capable to operate at a measured peak efficiency up to 3.08 Gop/s/mW at 1.24 mW peak power. To implement big RNN models without incurring in huge memory transfer overhead, multiple Chipmunk engines can cooperate to form a single systolic array. In this way, the Chipmunk architecture in a 75 tiles configuration can achieve real-time phoneme extraction on a demanding RNN topology proposed by Graves et al., consuming less than 13 mW of average power.