Marian Verhelst

Yasmine Omri, Ziyu Gan, Zachary Broveak et al.

LLM agents are increasingly deployed on long-horizon tasks requiring sustained reasoning over extended interaction histories. Realizing this at scale requires agents to persistently store, retrieve, and update their own memory across sessions. A rich ecosystem of agent memory systems has emerged spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows. Yet, their system-level behavior remains uncharacterized. We present the first systems characterization of agent memory. First, we introduce a system-oriented taxonomy classifying agent memory systems along four axes. Second, we build a phase-aware profiling harness attributing cost to construction, retrieval, and generation. Third, we characterize ten representative systems across two benchmark suites, uncovering how design choices shift cost across the write and read paths. Finally, we derive 10 system recommendations covering construction scheduling, capability floors, amortization via query volume, freshness-latency tradeoffs, and fleet-scale management.

Matteo Risso, Alessio Burrello, Giuseppe Maria Sarda et al.

The need to execute Deep Neural Networks (DNNs) at low latency and low power at the edge has spurred the development of new heterogeneous Systems-on-Chips (SoCs) encapsulating a diverse set of hardware accelerators. How to optimally map a DNN onto such multi-accelerator systems is an open problem. We propose ODiMO, a hardware-aware tool that performs a fine-grain mapping across different accelerators on-chip, splitting individual layers and executing them in parallel, to reduce inference energy consumption or latency, while taking into account each accelerator's quantization precision to maintain accuracy. Pareto-optimal networks in the accuracy vs. energy or latency space are pursued for three popular dataset/DNN pairs, and deployed on the DIANA heterogeneous ultra-low power edge AI SoC. We show that ODiMO reduces energy/latency by up to 33%/31% with limited accuracy drop (-0.53%/-0.32%) compared to manual heuristic mappings.

Victor J. B. Jung, Arne Symons, Linyan Mei et al.

To meet the growing need for computational power for DNNs, multiple specialized hardware architectures have been proposed. Each DNN layer should be mapped onto the hardware with the most efficient schedule, however, SotA schedulers struggle to consistently provide optimum schedules in a reasonable time across all DNN-HW combinations. This paper proposes SALSA, a fast dual-engine scheduler to generate optimal execution schedules for both even and uneven mapping. We introduce a new strategy, combining exhaustive search with simulated annealing to address the dynamic nature of the loop ordering design space size across layers. SALSA is extensively benchmarked against two SotA schedulers, LOMA and Timeloop on 5 different DNNs, on average SALSA finds schedules with 11.9% and 7.6% lower energy while speeding up the search by 1.7x and 24x compared to LOMA and Timeloop, respectively.

Jason Yik, Korneel Van den Berghe, Douwe den Blanken et al. · eth-zurich

Neuromorphic computing shows promise for advancing computing efficiency and capabilities of AI applications using brain-inspired principles. However, the neuromorphic research field currently lacks standardized benchmarks, making it difficult to accurately measure technological advancements, compare performance with conventional methods, and identify promising future research directions. Prior neuromorphic computing benchmark efforts have not seen widespread adoption due to a lack of inclusive, actionable, and iterative benchmark design and guidelines. To address these shortcomings, we present NeuroBench: a benchmark framework for neuromorphic computing algorithms and systems. NeuroBench is a collaboratively-designed effort from an open community of researchers across industry and academia, aiming to provide a representative structure for standardizing the evaluation of neuromorphic approaches. The NeuroBench framework introduces a common set of tools and systematic methodology for inclusive benchmark measurement, delivering an objective reference framework for quantifying neuromorphic approaches in both hardware-independent (algorithm track) and hardware-dependent (system track) settings. In this article, we outline tasks and guidelines for benchmarks across multiple application domains, and present initial performance baselines across neuromorphic and conventional approaches for both benchmark tracks. NeuroBench is intended to continually expand its benchmarks and features to foster and track the progress made by the research community.

Robin Geens, Joran Heldens, Joren Dumoulin et al.

Ternary weight quantization (e.g., BitNet b1.58) offers a promising path to mitigate the memory bandwidth bottleneck in Large Language Model (LLM) inference. However, conventional compute platforms lack native support for ternary-weight arithmetic, often relying on inefficient dequantization. Lookup table (LUT)-based hardware architectures provide an effective alternative by replacing multiplications with conditional additions, but their design space remains largely unexplored. Existing designs rely on heuristic parameter selection, lacking a systematic understanding of the architectural trade-offs. This work addresses this gap by formalizing the design space of ternary LUT-based accelerators and presenting an open-source hardware generator coupled with an analytical cost model, validated against synthesis in TSMC 16nm technology. By spanning the full architectural space, this framework not only enables rapid design space exploration but also establishes a common footing for fair cross-design evaluation, which was previously hindered by inconsistent instantiations across published accelerators. Using this framework, we challenge several assumptions and design choices in recent literature. We demonstrate that the optimal architecture is fundamentally governed by the activation data type: while LUT-based reuse offers significant gains for high-cost arithmetic (e.g., FP16), it yields diminishing returns for small integer types. Furthermore, we show that maximizing core size consistently improves area density compared to highly tiled approaches. Our optimized designs achieve a 2.2x area reduction compared to multiplier-based baselines. Moreover, by benchmarking state-of-the-art implementations against our model, we reveal that correcting suboptimal parameters yields up to a 1.2x area improvement.

Zuzana Jelčicová, Marian Verhelst

Multi-head self-attention forms the core of Transformer networks. However, their quadratically growing complexity with respect to the input sequence length impedes their deployment on resource-constrained edge devices. We address this challenge by proposing a dynamic pruning method, which exploits the temporal stability of data across tokens to reduce inference cost. The threshold-based method only retains significant differences between the subsequent tokens, effectively reducing the number of multiply-accumulates, as well as the internal tensor data sizes. The approach is evaluated on the Google Speech Commands Dataset for keyword spotting, and the performance is compared against the baseline Keyword Transformer. Our experiments show that we can reduce ~80% of operations while maintaining the original 98.4% accuracy. Moreover, a reduction of ~87-94% operations can be achieved when only degrading the accuracy by 1-4%, speeding up the multi-head self-attention inference by a factor of ~7.5-16.

Yuzong Chen, Chao Fang, Xilai Dai et al.

The substantial memory bandwidth and computational demands of large language models (LLMs) present critical challenges for efficient inference. To tackle this, the literature has explored heterogeneous systems that combine neural processing units (NPUs) with DRAM-based processing-in-memory (PIM) for LLM acceleration. However, existing high-precision (e.g., FP16) PIM compute units incur significant area and power overhead in DRAM technology, limiting the effective computation throughput. In this paper, we introduce P3-LLM, a novel NPU-PIM integrated accelerator for LLM inference using hybrid numerical formats. Our approach is threefold: First, we propose a flexible mixed-precision quantization scheme, which leverages hybrid numerical formats to quantize different LLM operands with high compression efficiency and minimal accuracy loss. Second, we architect an efficient PIM accelerator for P3-LLM, featuring enhanced compute units to support hybrid numerical formats. Our careful choice of numerical formats allows to co-design low-precision PIM compute units that significantly boost the computation throughput under iso-area constraints. Third, we optimize the low-precision dataflow of different LLM modules by applying operator fusion to minimize the overhead of runtime dequantization. Evaluation on a diverse set of representative LLMs and tasks demonstrates that P3-LLM achieves state-of-the-art accuracy in terms of both KV-cache quantization and weight-activation quantization. Combining the proposed quantization scheme with PIM architecture co-design, P3-LLM yields an average of $4.9\times$, $2.0\times$, and $3.4\times$ speedups over the state-of-the-art LLM accelerators HBM-PIM, Ecco, and Pimba, respectively. Our quantization code is available at https://github.com/yc2367/P3-LLM.git

Maxim Bonnaerens, Matthias Freiberger, Marian Verhelst et al.

In this work we propose a methodology to accurately evaluate and compare the performance of efficient neural network building blocks for computer vision in a hardware-aware manner. Our comparison uses pareto fronts based on randomly sampled networks from a design space to capture the underlying accuracy/complexity trade-offs. We show that our approach allows to match the information obtained by previous comparison paradigms, but provides more insights in the relationship between hardware cost and accuracy. We use our methodology to analyze different building blocks and evaluate their performance on a range of embedded hardware platforms. This highlights the importance of benchmarking building blocks as a preselection step in the design process of a neural network. We show that choosing the right building block can speed up inference by up to a factor of 2x on specific hardware ML accelerators.

Xiaoling Yi, Ryan Antonio, Joren Dumoulin et al.

Deep neural networks (DNNs) face significant challenges when deployed on resource-constrained extreme edge devices due to their computational and data-intensive nature. While standalone accelerators tailored for specific application scenarios suffer from inflexible control and limited programmability, generic hardware acceleration platforms coupled with RISC-V CPUs can enable high reusability and flexibility, yet typically at the expense of system level efficiency and low utilization. To fill this gap, we propose OpenGeMM, an open-source acceleration platform, jointly demonstrating high efficiency and utilization, as well as ease of configurability and programmability. OpenGeMM encompasses a parameterized Chisel-coded GeMM accelerator, a lightweight RISC-V processor, and a tightly coupled multi-banked scratchpad memory. The GeMM core utilization and system efficiency are boosted through three mechanisms: configuration pre-loading, input pre-fetching with output buffering, and programmable strided memory access. Experimental results show that OpenGeMM can consistently achieve hardware utilization ranging from 81.89% to 99.34% across diverse CNN and Transformer workloads. Compared to the SotA open-source Gemmini accelerator, OpenGeMM demonstrates a 3.58x to 16.40x speedup on normalized throughput across a wide variety ofGeMM workloads, while achieving 4.68 TOPS/W system efficiency.

Stef Cuyckens, Xiaoling Yi, Robin Geens et al.

Emerging continual learning applications necessitate next-generation neural processing unit (NPU) platforms to support both training and inference operations. The promising Microscaling (MX) standard enables narrow bit-widths for inference and large dynamic ranges for training. However, existing MX multiply-accumulate (MAC) designs face a critical trade-off: integer accumulation requires expensive conversions from narrow floating-point products, while FP32 accumulation suffers from quantization losses and costly normalization. To address these limitations, we propose a hybrid precision-scalable reduction tree for MX MACs that combines the benefits of both approaches, enabling efficient mixed-precision accumulation with controlled accuracy relaxation. Moreover, we integrate an 8x8 array of these MACs into the state-of-the-art (SotA) NPU integration platform, SNAX, to provide efficient control and data transfer to our optimized precision-scalable MX datapath. We evaluate our design both on MAC and system level and compare it to the SotA. Our integrated system achieves an energy efficiency of 657, 1438-1675, and 4065 GOPS/W, respectively, for MXINT8, MXFP8/6, and MXFP4, with a throughput of 64, 256, and 512 GOPS.

Robin Geens, Arne Symons, Marian Verhelst

State Space Models (SSMs) offer a promising alternative to transformers for long-sequence processing. However, their efficiency remains hindered by memory-bound operations, particularly in the prefill stage. While MARCA, a recent first effort to accelerate SSMs through a dedicated hardware accelerator, achieves great speedup over high-end GPUs, an analysis into the broader accelerator design space is lacking. This work systematically analyzes SSM acceleration opportunities both from the scheduling perspective through fine-grained operator fusion and the hardware perspective through design space exploration, using an extended version of the Stream modeling framework. Our results demonstrate that the improved data locality stemming from our optimized fusion and scheduling strategy enables a speedup of up to 4.8x over unfused execution, while our adaptive memory-aware fusion approach reduces on-chip memory requirements by an order of magnitude without sacrificing performance. We further explore accelerator design trade-offs, showing that a fusion-aware hardware architecture can achieve 1.78x higher performance than the state-of-the-art MARCA accelerator, within the same area budget. These results establish operator fusion as a key enabler for next-generation SSM accelerators.

Robin Geens, Marian Verhelst

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, improves the efficiency of large language models by projecting query, key, and value tensors into a compact latent space. This architectural change reduces the KV-cache size and significantly lowers memory bandwidth demands, particularly in the autoregressive decode phase. This letter presents the first hardware-centric analysis of MLA, comparing it to conventional Multi-Head Attention (MHA) and evaluating its implications for accelerator performance. We identify two alternative execution schemes of MLA--reusing, resp. recomputing latent projection matrices--which offer distinct trade-offs between compute and memory access. Using the Stream design space exploration framework, we model their throughput and energy cost across a range of hardware platforms and find that MLA can shift attention workloads toward the compute-bound regime. Our results show that MLA not only reduces bandwidth usage but also enables adaptable execution strategies aligned with hardware constraints. Compared to MHA, it provides more stable and efficient performance, particularly on bandwidth-limited hardware platforms. These findings emphasize MLA's relevance as a co-design opportunity for future AI accelerators.

Jérémy Morlier, Robin Geens, Stef Cuyckens et al.

While hardware-software co-design has significantly improved the efficiency of neural network inference, modeling the training phase remains a critical yet underexplored challenge. Training workloads impose distinct constraints, particularly regarding memory footprint and backpropagation complexity, which existing inference-focused tools fail to capture. This paper introduces MONET, a framework designed to model the training of neural networks on heterogeneous dataflow accelerators. MONET builds upon Stream, an experimentally verified framework that that models the inference of neural networks on heterogeneous dataflow accelerators with layer fusion. Using MONET, we explore the design space of ResNet-18 and a small GPT-2, demonstrating the framework's capability to model training workflows and find better hardware architectures. We then further examine problems that become more complex in neural network training due to the larger design space, such as determining the best layer-fusion configuration. Additionally, we use our framework to find interesting trade-offs in activation checkpointing, with the help of a genetic algorithm. Our findings highlight the importance of a holistic approach to hardware-software co-design for scalable and efficient deep learning deployment.

Ryan Albert Antonio, Joren Dumoulin, Xiaoling Yi et al.

Heterogeneous accelerator-centric compute clusters are emerging as efficient solutions for diverse AI workloads. However, current integration strategies often compromise data movement efficiency and encounter compatibility issues in hardware and software. This prevents a unified approach that balances performance and ease of use. To this end, we present SNAX, an open-source integrated HW-SW framework enabling efficient multi-accelerator platforms through a novel hybrid-coupling scheme, consisting of loosely coupled asynchronous control and tightly coupled data access. SNAX brings reusable hardware modules designed to enhance compute accelerator utilization, and its customizable MLIR-based compiler to automate key system management tasks, jointly enabling rapid development and deployment of customized multi-accelerator compute clusters. Through extensive experimentation, we demonstrate SNAX's efficiency and flexibility in a low-power heterogeneous SoC. Accelerators can easily be integrated and programmed to achieve > 10x improvement in neural network performance compared to other accelerator systems while maintaining accelerator utilization of > 90% in full system operation.

Alaa Zniber, Arne Symons, Ouassim Karrakchou et al.

Advancements in high-performance computing and cloud technologies have enabled the development of increasingly sophisticated Deep Learning (DL) models. However, the growing demand for embedded intelligence at the edge imposes stringent computational and energy constraints, challenging the deployment of these large-scale models. Early Exiting Neural Networks (EENN) have emerged as a promising solution, allowing dynamic termination of inference based on input complexity to enhance efficiency. Despite their potential, EENN performance is highly influenced by the heterogeneity of edge accelerators and the constraints imposed by quantization, affecting accuracy, energy efficiency, and latency. Yet, research on the automatic optimization of EENN design for edge hardware remains limited. To bridge this gap, we propose a hardware-aware Neural Architecture Search (NAS) framework that systematically integrates the effects of quantization and hardware resource allocation to optimize the placement of early exit points within a network backbone. Experimental results on the CIFAR-10 dataset demonstrate that our NAS framework can discover architectures that achieve over a 50\% reduction in computational costs compared to conventional static networks, making them more suitable for deployment in resource-constrained edge environments.

Chao Fang, Man Shi, Robin Geens et al.

The widely-used, weight-only quantized large language models (LLMs), which leverage low-bit integer (INT) weights and retain floating-point (FP) activations, reduce storage requirements while maintaining accuracy. However, this shifts the energy and latency bottlenecks towards the FP activations that are associated with costly memory accesses and computations. Existing LLM accelerators focus primarily on computation optimizations, overlooking the potential of jointly optimizing FP computations and data movement, particularly for the dominant FP-INT GeMM operations in LLM inference. To address these challenges, we investigate the sensitivity of activation precision across various LLM modules and its impact on overall model accuracy. Based on our findings, we first propose the Anda data type: an adaptive data format with group-shared exponent bits and dynamic mantissa bit allocation. Secondly, we develop an iterative post-training adaptive precision search algorithm that optimizes the bit-width for different LLM modules to balance model accuracy, energy efficiency, and inference speed. Lastly, a suite of hardware optimization techniques is proposed to maximally exploit the benefits of the Anda format. These include a bit-plane-based data organization scheme, Anda-enhanced processing units with bit-serial computation, and a runtime bit-plane Anda compressor to simultaneously optimize storage, computation, and memory footprints. Our evaluations on FPINT GeMM operations show that Anda achieves a 2.4x speedup, 4.0x area efficiency, and 3.1x energy efficiency improvement on average for popular LLMs including OPT, LLaMA, and LLaMA-2 series over the GPU-like FP-FP baseline. Anda demonstrates strong adaptability across various application scenarios, accuracy requirements, and system performance, enabling efficient LLM inference across a wide range of deployment scenarios.

Mohamed Amine Hamdi, Francesco Daghero, Giuseppe Maria Sarda et al.

Streamlining the deployment of Deep Neural Networks (DNNs) on heterogeneous edge platforms, coupling within the same micro-controller unit (MCU) instruction processors and hardware accelerators for tensor computations, is becoming one of the crucial challenges of the TinyML field. The best-performing DNN compilation toolchains are usually deeply customized for a single MCU family, and porting to a different heterogeneous MCU family implies labor-intensive re-development of almost the entire compiler. On the opposite side, retargetable toolchains, such as TVM, fail to exploit the capabilities of custom accelerators, resulting in the generation of general but unoptimized code. To overcome this duality, we introduce MATCH, a novel TVM-based DNN deployment framework designed for easy agile retargeting across different MCU processors and accelerators, thanks to a customizable model-based hardware abstraction. We show that a general and retargetable mapping framework enhanced with hardware cost models can compete with and even outperform custom toolchains on diverse targets while only needing the definition of an abstract hardware model and a SoC-specific API. We tested MATCH on two state-of-the-art heterogeneous MCUs, GAP9 and DIANA. On the four DNN models of the MLPerf Tiny suite MATCH reduces inference latency by up to 60.88 times on DIANA, compared to using the plain TVM, thanks to the exploitation of the on-board HW accelerator. Compared to HTVM, a fully customized toolchain for DIANA, we still reduce the latency by 16.94%. On GAP9, using the same benchmarks, we improve the latency by 2.15 times compared to the dedicated DORY compiler, thanks to our heterogeneous DNN mapping approach that synergically exploits the DNN accelerator and the eight-cores cluster available on board.

Robin Geens, Jonas De Schouwer, Marian Verhelst et al.

The Hardware Lottery posits that research directions are dictated by available silicon compute platforms. We identify a derivative phenomenon, the Hyperscale Lottery, where model architectures are optimized for cloud throughput at the expense of algorithmic efficiency. While State-Space Models (SSMs) such as Mamba were lauded for their linear complexity, ideal for edge intelligence, their evolution from Mamba-1 to Mamba-3 reveals a systematic divergence from edge-native efficiency. We demonstrate that Mamba-3's architectural changes, designed to saturate hyperscale GPUs, impose a significant edge penalty: a 28% latency increase at 880M parameters, worsening to 48% for 15M-parameter models. We argue for decoupling cloud-scale saturation strategies from core architectural design to preserve the viability of single-user, real-time edge intelligence.

Stef Cuyckens, Ryan Antonio, Chao Fang et al.

Implantable devices for reliable intracranial electroencephalography (iEEG) require efficient, accurate, and real-time detection of seizures. Dense hyperdimensional computing (HDC) proves to be efficient over neural networks; however, it still consumes considerable switching power for an ultra-low energy application. Sparse HDC, on the other hand, has the potential of further reducing the energy consumption, yet at the expense of having to support more complex operations and introducing an extra hyperparameter, the maximum hypervector density. To improve the energy and area efficiency of the sparse HDC operations, this work introduces the compressed item memory (CompIM) and simplifies the spatial bundling. We also analyze how a proper hyperparameter choice improves the detection delay compared to dense HDC. Ultimately, our optimizations achieve a 1.73x more energy- and 2.20x more area-efficient hardware design than the naive sparse implementation. We are also 7.50x more energy- and 3.24x more area-efficient than the dense HDC implementation. This work highlights the hardware advantages of sparse HDC, demonstrating its potential to enable smaller brain implants with a substantially extended battery life compared to the current state-of-the-art.

Shirui Zhao, Jun Yin, Lingyun Yao et al.

An increasing number of applications are exploiting sampling-based algorithms for planning, optimization, and inference. The Markov Chain Monte Carlo (MCMC) algorithms form the computational backbone of this emerging branch of machine learning. Unfortunately, the high computational cost limits their feasibility for large-scale problems and real-world applications, and the existing MCMC acceleration solutions are either limited in hardware flexibility or fail to maintain efficiency at the system level across a variety of end-to-end applications. This paper introduces \textbf{MC$^2$A}, an algorithm-hardware co-design framework, enabling efficient and flexible optimization for MCMC acceleration. Firstly, \textbf{MC$^2$A} analyzes the MCMC workload diversity through an extension of the processor performance roofline model with a 3rd dimension to derive the optimal balance between the compute, sampling and memory parameters. Secondly, \textbf{MC$^2$A} proposes a parametrized hardware accelerator architecture with flexible and efficient support of MCMC kernels with a pipeline of ISA-programmable tree-structured processing units, reconfigurable samplers and a crossbar interconnect to support irregular access. Thirdly, the core of \textbf{MC$^2$A} is powered by a novel Gumbel sampler that eliminates exponential and normalization operations. In the end-to-end case study, \textbf{MC$^2$A} achieves an overall {$307.6\times$, $1.4\times$, $2.0\times$, $84.2\times$} speedup compared to the CPU, GPU, TPU and state-of-the-art MCMC accelerator. Evaluated on various representative MCMC workloads, this work demonstrates and exploits the feasibility of general hardware acceleration to popularize MCMC-based solutions in diverse application domains.

Nimish Shah, Laura I. Galindez Olascoaga, Wannes Meert et al.

Bayesian reasoning is a powerful mechanism for probabilistic inference in smart edge-devices. During such inferences, a low-precision arithmetic representation can enable improved energy efficiency. However, its impact on inference accuracy is not yet understood. Furthermore, general-purpose hardware does not natively support low-precision representation. To address this, we propose ProbLP, a framework that automates the analysis and design of low-precision probabilistic inference hardware. It automatically chooses an appropriate energy-efficient representation based on worst-case error-bounds and hardware energy-models. It generates custom hardware for the resulting inference network exploiting parallelism, pipelining and low-precision operation. The framework is validated on several embedded-sensing benchmarks.

Robby Neven, Marian Verhelst, Tinne Tuytelaars et al.

Training deep learning models on embedded devices is typically avoided since this requires more memory, computation and power over inference. In this work, we focus on lowering the amount of memory needed for storing all activations, which are required during the backward pass to compute the gradients. Instead, during the forward pass, static Synthetic Gradient Modules (SGMs) predict gradients for each layer. This allows training the model in a feed-forward manner without having to store all activations. We tested our method on a robot grasping scenario where a robot needs to learn to grasp new objects given only a single demonstration. By first training the SGMs in a meta-learning manner on a set of common objects, during fine-tuning, the SGMs provided the model with accurate gradients to successfully learn to grasp new objects. We have shown that our method has comparable results to using standard backpropagation.

Colby R. Banbury, Vijay Janapa Reddi, Max Lam et al.

Recent advancements in ultra-low-power machine learning (TinyML) hardware promises to unlock an entirely new class of smart applications. However, continued progress is limited by the lack of a widely accepted benchmark for these systems. Benchmarking allows us to measure and thereby systematically compare, evaluate, and improve the performance of systems and is therefore fundamental to a field reaching maturity. In this position paper, we present the current landscape of TinyML and discuss the challenges and direction towards developing a fair and useful hardware benchmark for TinyML workloads. Furthermore, we present our four benchmarks and discuss our selection methodology. Our viewpoints reflect the collective thoughts of the TinyMLPerf working group that is comprised of over 30 organizations.

Gert Dekkers, Fernando Rosas, Steven Lauwereins et al.

Smart sensing is expected to become a pervasive technology in smart cities and environments of the near future. These services are improving their capabilities due to integrated devices shrinking in size while maintaining their computational power, which can run diverse Machine Learning algorithms and achieve high performance in various data-processing tasks. One attractive sensor modality to be used for smart sensing are acoustic sensors, which can convey highly informative data while keeping a moderate energy consumption. Unfortunately, the energy budget of current wireless sensor networks is usually not enough to support the requirements of standard microphones. Therefore, energy efficiency needs to be increased at all layers --- sensing, signal processing and communication --- in order to bring wireless smart acoustic sensors into the market. To help to attain this goal, this paper introduces WASN-EM: an energy consumption model for wireless acoustic sensors networks (WASN), whose aim is to aid in the development of novel techniques to increase the energy-efficient of smart wireless acoustic sensors. This model provides a first step of exploration prior to custom design of a smart wireless acoustic sensor, and also can be used to compare the energy consumption of different protocols.

Bert Moons, Daniel Bankman, Lita Yang et al.

This paper introduces BinarEye: a digital processor for always-on Binary Convolutional Neural Networks. The chip maximizes data reuse through a Neuron Array exploiting local weight Flip-Flops. It stores full network models and feature maps and hence requires no off-chip bandwidth, which leads to a 230 1b-TOPS/W peak efficiency. Its 3 levels of flexibility - (a) weight reconfiguration, (b) a programmable network depth and (c) a programmable network width - allow trading energy for accuracy depending on the task's requirements. BinarEye's full system input-to-label energy consumption ranges from 14.4uJ/f for 86% CIFAR-10 and 98% owner recognition down to 0.92uJ/f for 94% face detection at up to 1700 frames per second. This is 3-12-70x more efficient than the state-of-the-art at on-par accuracy.

Matthijs Van keirsbilck, Bert Moons, Marian Verhelst

Today's Automatic Speech Recognition systems only rely on acoustic signals and often don't perform well under noisy conditions. Performing multi-modal speech recognition - processing acoustic speech signals and lip-reading video simultaneously - significantly enhances the performance of such systems, especially in noisy environments. This work presents the design of such an audio-visual system for Automated Speech Recognition, taking memory and computation requirements into account. First, a Long-Short-Term-Memory neural network for acoustic speech recognition is designed. Second, Convolutional Neural Networks are used to model lip-reading features. These are combined with an LSTM network to model temporal dependencies and perform automatic lip-reading on video. Finally, acoustic-speech and visual lip-reading networks are combined to process acoustic and visual features simultaneously. An attention mechanism ensures performance of the model in noisy environments. This system is evaluated on the TCD-TIMIT 'lipspeaker' dataset for audio-visual phoneme recognition with clean audio and with additive white noise at an SNR of 0dB. It achieves 75.70% and 58.55% phoneme accuracy respectively, over 14 percentage points better than the state-of-the-art for all noise levels.

Bert Moons, Koen Goetschalckx, Nick Van Berckelaer et al.

This work targets the automated minimum-energy optimization of Quantized Neural Networks (QNNs) - networks using low precision weights and activations. These networks are trained from scratch at an arbitrary fixed point precision. At iso-accuracy, QNNs using fewer bits require deeper and wider network architectures than networks using higher precision operators, while they require less complex arithmetic and less bits per weights. This fundamental trade-off is analyzed and quantified to find the minimum energy QNN for any benchmark and hence optimize energy-efficiency. To this end, the energy consumption of inference is modeled for a generic hardware platform. This allows drawing several conclusions across different benchmarks. First, energy consumption varies orders of magnitude at iso-accuracy depending on the number of bits used in the QNN. Second, in a typical system, BinaryNets or int4 implementations lead to the minimum energy solution, outperforming int8 networks up to 2-10x at iso-accuracy. All code used for QNN training is available from https://github.com/BertMoons.

Bert Moons, Bert De Brabandere, Luc Van Gool et al.

Recently ConvNets or convolutional neural networks (CNN) have come up as state-of-the-art classification and detection algorithms, achieving near-human performance in visual detection. However, ConvNet algorithms are typically very computation and memory intensive. In order to be able to embed ConvNet-based classification into wearable platforms and embedded systems such as smartphones or ubiquitous electronics for the internet-of-things, their energy consumption should be reduced drastically. This paper proposes methods based on approximate computing to reduce energy consumption in state-of-the-art ConvNet accelerators. By combining techniques both at the system- and circuit level, we can gain energy in the systems arithmetic: up to 30x without losing classification accuracy and more than 100x at 99% classification accuracy, compared to the commonly used 16-bit fixed point number format.