Lukas Cavigelli

Lorenz K. Muller, Philippe Bich, Chiara Boretti et al.

Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN

Renzo Andri, Beatrice Bussolino, Antonio Cipolletta et al.

Most of today's computer vision pipelines are built around deep neural networks, where convolution operations require most of the generally high compute effort. The Winograd convolution algorithm computes convolutions with fewer MACs compared to the standard algorithm, reducing the operation count by a factor of 2.25x for 3x3 convolutions when using the version with 2x2-sized tiles $F_2$. Even though the gain is significant, the Winograd algorithm with larger tile sizes, i.e., $F_4$, offers even more potential in improving throughput and energy efficiency, as it reduces the required MACs by 4x. Unfortunately, the Winograd algorithm with larger tile sizes introduces numerical issues that prevent its use on integer domain-specific accelerators and higher computational overhead to transform input and output data between spatial and Winograd domains. To unlock the full potential of Winograd $F_4$, we propose a novel tap-wise quantization method that overcomes the numerical issues of using larger tiles, enabling integer-only inference. Moreover, we present custom hardware units that process the Winograd transformations in a power- and area-efficient way, and we show how to integrate such custom modules in an industrial-grade, programmable DSA. An extensive experimental evaluation on a large set of state-of-the-art computer vision benchmarks reveals that the tap-wise quantization algorithm makes the quantized Winograd $F_4$ network almost as accurate as the FP32 baseline. The Winograd-enhanced DSA achieves up to 1.85x gain in energy efficiency and up to 1.83x end-to-end speed-up for state-of-the-art segmentation and detection networks.

Antoine Scardigli, Lukas Cavigelli, Lorenz K. Müller

Monte-Carlo path tracing is a powerful technique for realistic image synthesis but suffers from high levels of noise at low sample counts, limiting its use in real-time applications. To address this, we propose a framework with end-to-end training of a sampling importance network, a latent space encoder network, and a denoiser network. Our approach uses reinforcement learning to optimize the sampling importance network, thus avoiding explicit numerically approximated gradients. Our method does not aggregate the sampled values per pixel by averaging but keeps all sampled values which are then fed into the latent space encoder. The encoder replaces handcrafted spatiotemporal heuristics by learned representations in a latent space. Finally, a neural denoiser is trained to refine the output image. Our approach increases visual quality on several challenging datasets and reduces rendering times for equal quality by a factor of 1.6x compared to the previous state-of-the-art, making it a promising solution for real-time applications.

Jannis Schönleber, Lukas Cavigelli, Matteo Perotti et al.

Artificial intelligence has surged in recent years, with advancements in machine learning rapidly impacting nearly every area of life. However, the growing complexity of these models has far outpaced advancements in available hardware accelerators, leading to significant computational and energy demands, primarily due to matrix multiplications, which dominate the compute workload. Maddness (i.e., Multiply-ADDitioN-lESS) presents a hash-based version of product quantization, which renders matrix multiplications into lookups and additions, eliminating the need for multipliers entirely. We present Stella Nera, the first Maddness-based accelerator achieving an energy efficiency of 161 TOp/s/W@0.55V, 25x better than conventional MatMul accelerators due to its small components and reduced computational complexity. We further enhance Maddness with a differentiable approximation, allowing for gradient-based fine-tuning and achieving an end-to-end performance of 92.5% Top-1 accuracy on CIFAR-10.

Luca Benfenati, Matteo Risso, Andrea Vannozzi et al.

Key--value (KV) caching enables fast autoregressive decoding but at long contexts becomes a dominant bottleneck in High Bandwidth Memory (HBM) capacity and bandwidth. A common mitigation is to compress cached keys and values by projecting per-head matrixes to a lower rank, storing only the projections in the HBM. However, existing post-training approaches typically fit these projections using SVD-style proxy objectives, which may poorly reflect end-to-end reconstruction after softmax, value mixing, and subsequent decoder-layer transformations. For these reasons, we introduce StiefAttention, a post-training KV-cache compression method that learns \emph{orthonormal} projection bases by directly minimizing \emph{decoder-layer output reconstruction error}. StiefAttention additionally precomputes, for each layer, an error-rank profile over candidate ranks, enabling flexible layer-wise rank allocation under a user-specified error budget. Noteworthy, on Llama3-8B under the same conditions, StiefAttention outperforms EigenAttention by $11.9$ points on C4 perplexity and $5.4\%$ on 0-shot MMLU accuracy at iso-compression, yielding lower relative error and higher cosine similarity with respect to the original decoder-layer outputs.

Ahmet Caner Yüzügüler, Jiawei Zhuang, Lukas Cavigelli

Large language models (LLMs) are typically served from clusters of GPUs/NPUs that consist of large number of devices. Unfortunately, communication between these devices incurs significant overhead, increasing the inference latency and cost while limiting the scalability. Prior work addressed this issue by overlapping communication with compute, but has severe limitations due to the data dependencies between these operations. In this paper, we propose PRESERVE, a novel framework that prefetches model weights and KV-cache from off-chip HBM memory to the on-chip cache of AI accelerators during the communication operations, which offers various advantages and performance improvements compared to prior methods. Through extensive experiments conducted on commercial AI accelerators, we demonstrate up to 1.6x end-to-end speedup on state-of-the-art, open-source LLMs. Additionally, we perform a design space exploration that identifies the optimal hardware configuration for the proposed method, showing a further 1.25x improvement in performance per cost by selecting the optimal L2 cache size. Our results show that PRESERVE has the potential to mitigate the memory bottlenecks and communication overheads, offering a solution to improve the performance and scalability of the LLM inference systems.

Lorenz K. Müller, Philippe Bich, Jiawei Zhuang et al.

Post-training quantization has emerged as the most widely used strategy for deploying large language models at low precision. Still, current methods show perplexity degradation at bit-widths less than or equal to 4, partly because representing outliers causes precision issues in parameters that share the same scales as these outliers. This problem is especially pronounced for calibration-free, uniform quantization methods. We introduce SINQ to augment existing post-training quantizers with an additional second-axis scale factor and a fast Sinkhorn-Knopp-style algorithm that finds scales to normalize per-row and per-column variances, thereby minimizing a novel per-matrix proxy target for quantization: the matrix imbalance. Our method has no interactions between layers and can be trivially applied to new architectures to quantize any linear layers. We evaluate our method on the Qwen3 model family and DeepSeek-V2.5. SINQ improves WikiText2 and C4 perplexity significantly against uncalibrated uniform quantization baselines and can be further enhanced by combining it with calibration and non-uniform quantization levels. Code to reproduce the results of this work and to easily quantize models using SINQ is available at https://github.com/huawei-csl/SINQ.

Xiaying Wang, Michele Magno, Lukas Cavigelli et al.

The growing number of low-power smart devices in the Internet of Things is coupled with the concept of "Edge Computing", that is moving some of the intelligence, especially machine learning, towards the edge of the network. Enabling machine learning algorithms to run on resource-constrained hardware, typically on low-power smart devices, is challenging in terms of hardware (optimized and energy-efficient integrated circuits), algorithmic and firmware implementations. This paper presents FANN-on-MCU, an open-source toolkit built upon the Fast Artificial Neural Network (FANN) library to run lightweight and energy-efficient neural networks on microcontrollers based on both the ARM Cortex-M series and the novel RISC-V-based Parallel Ultra-Low-Power (PULP) platform. The toolkit takes multi-layer perceptrons trained with FANN and generates code targeted at execution on low-power microcontrollers either with a floating-point unit (i.e., ARM Cortex-M4F and M7F) or without (i.e., ARM Cortex M0-M3 or PULP-based processors). This paper also provides an architectural performance evaluation of neural networks on the most popular ARM Cortex-M family and the parallel RISC-V processor called Mr. Wolf. The evaluation includes experimental results for three different applications using a self-sustainable wearable multi-sensor bracelet. Experimental results show a measured latency in the order of only a few microseconds and a power consumption of few milliwatts while keeping the memory requirements below the limitations of the targeted microcontrollers. In particular, the parallel implementation on the octa-core RISC-V platform reaches a speedup of 22x and a 69% reduction in energy consumption with respect to a single-core implementation on Cortex-M4 for continuous real-time classification.

Cristian Cioflan, Lukas Cavigelli, Manuele Rusci et al.

Keyword spotting accuracy degrades when neural networks are exposed to noisy environments. On-site adaptation to previously unseen noise is crucial to recovering accuracy loss, and on-device learning is required to ensure that the adaptation process happens entirely on the edge device. In this work, we propose a fully on-device domain adaptation system achieving up to 14% accuracy gains over already-robust keyword spotting models. We enable on-device learning with less than 10 kB of memory, using only 100 labeled utterances to recover 5% accuracy after adapting to the complex speech noise. We demonstrate that domain adaptation can be achieved on ultra-low-power microcontrollers with as little as 806 mJ in only 14 s on always-on, battery-operated devices.

Cristian Cioflan, Lukas Cavigelli, Luca Benini

Keyword spotting systems for always-on TinyML-constrained applications require on-site tuning to boost the accuracy of offline trained classifiers when deployed in unseen inference conditions. Adapting to the speech peculiarities of target users requires many in-domain samples, often unavailable in real-world scenarios. Furthermore, current on-device learning techniques rely on computationally intensive and memory-hungry backbone update schemes, unfit for always-on, battery-powered devices. In this work, we propose a novel on-device learning architecture, composed of a pretrained backbone and a user-aware embedding learning the user's speech characteristics. The so-generated features are fused and used to classify the input utterance. For domain shifts generated by unseen speakers, we measure error rate reductions of up to 19% from 30.1% to 24.3% based on the 35-class problem of the Google Speech Commands dataset, through the inexpensive update of the user projections. We moreover demonstrate the few-shot learning capabilities of our proposed architecture in sample- and class-scarce learning conditions. With 23.7 kparameters and 1 MFLOP per epoch required for on-device training, our system is feasible for TinyML applications aimed at battery-powered microcontrollers.

Konstantin Berestizshevsky, Renzo Andri, Lukas Cavigelli

We present Top-Theta (Top-$θ$) Attention, a training-free method for sparsifying transformer attention during inference. Our key insight is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row. This approach enables content-based sparsity without retraining, and it remains robust across data domains. We further introduce compensation techniques to preserve accuracy under aggressive sparsification, establishing attention thresholding as a practical and principled alternative to top-k attention. We provide extensive evaluation on natural language processing tasks, showing that Top-$θ$ achieves 3-10x reduction in V-cache usage and up to 10x fewer attention elements during inference while degrading no more than 1% in accuracy.

Michele Marzollo, Jiawei Zhuang, Niklas Roemer et al.

Over the past year, Speculative Decoding has gained popularity as a technique for accelerating Large Language Model inference. While several methods have been introduced, most struggle to deliver satisfactory performance at batch sizes typical for data centers ($\geq 8$) and often involve significant deployment complexities. In this work, we offer a theoretical explanation of how Speculative Decoding can be effectively utilized with larger batch sizes. We also introduce a method that integrates seamlessly into existing systems without additional training or the complexity of deploying a small LLM. In a continuous batching setting, we achieve a 4x increase in throughput without any latency impact for short context generation, and a 1.7-2x improvement in both latency and throughput for longer contexts.

Felix Arnold, Maxence Bouvier, Ryan Amaudruz et al.

This work introduces an innovative method for improving combinational digital circuits through random exploration in MIG-based synthesis. High-quality circuits are crucial for performance, power, and cost, making this a critical area of active research. Our approach incorporates next-state prediction and iterative selection, significantly accelerating the synthesis process. This novel method achieves up to 14x synthesis speedup and up to 20.94% better MIG minimization on the EPFL Combinational Benchmark Suite compared to state-of-the-art techniques. We further explore various predictor models and show that increased prediction accuracy does not guarantee an equivalent increase in synthesis quality of results or speedup, observing that randomness remains a desirable factor.

Vasilije Stambolic, Aritra Dhar, Lukas Cavigelli

Retrieval-Augmented Generation (RAG) increases the reliability and trustworthiness of the LLM response and reduces hallucination by eliminating the need for model retraining. It does so by adding external data into the LLM's context. We develop a new class of black-box attack, RAG-Pull, that inserts hidden UTF characters into queries or external code repositories, redirecting retrieval toward malicious code, thereby breaking the models' safety alignment. We observe that query and code perturbations alone can shift retrieval toward attacker-controlled snippets, while combined query-and-target perturbations achieve near-perfect success. Once retrieved, these snippets introduce exploitable vulnerabilities such as remote code execution and SQL injection. RAG-Pull's minimal perturbations can alter the model's safety alignment and increase preference towards unsafe code, therefore opening up a new class of attacks on LLMs.

Shai Bergman, Won Wook Song, Lukas Cavigelli et al.

Existing storage systems lack visibility into workload intent, limiting their ability to adapt to the semantics of modern, large-scale data-intensive applications. This disconnect leads to brittle heuristics and fragmented, siloed optimizations. To address these limitations, we propose Intent-Driven Storage Systems (IDSS), a vision for a new paradigm where large language models (LLMs) infer workload and system intent from unstructured signals to guide adaptive and cross-layer parameter reconfiguration. IDSS provides holistic reasoning for competing demands, synthesizing safe and efficient decisions within policy guardrails. We present four design principles for integrating LLMs into storage control loops and propose a corresponding system architecture. Initial results on FileBench workloads show that IDSS can improve IOPS by up to 2.45X by interpreting intent and generating actionable configurations for storage components such as caching and prefetching. These findings suggest that, when constrained by guardrails and embedded within structured workflows, LLMs can function as high-level semantic optimizers, bridging the gap between application goals and low-level system control. IDSS points toward a future in which storage systems are increasingly adaptive, autonomous, and aligned with dynamic workload demands.

Ahmet Caner Yüzügüler, Ahmet Çelik, Jiawei Zhuang et al.

Multi-Head Latent Attention (MLA) is a recent attention mechanism adopted in state-of-the-art LLMs such as DeepSeek-v3 and Kimi K2. Thanks to its novel formulation, MLA allows two functionally equivalent but computationally distinct kernel implementations: naive and absorb. While the naive kernels (e.g., FlashAttention) are typically preferred in training and prefill for their computational efficiency, existing decoding kernels (e.g., FlashMLA) rely on the absorb method to minimize HBM bandwidth usage. However, the compute-bound nature of the absorb implementations prohibits performance benefits from data reuse opportunities in attention calculations, such as shared prefixes. In this work, we introduce TyphoonMLA, a hybrid approach that combines naive and absorb formulations to harness the strengths of both. TyphoonMLA effectively leverages the shared prefix by applying the naive formulation to the compute-bound parts of attention calculations, while reducing the bandwidth requirements for non-shared parts by using the absorb formulation. As a result, TyphoonMLA improves the throughput of attention calculations in MLA architectures by up to 3x and 3.24x on NPU and GPUs, with only a 3% overhead in HBM size.

Maxence Bouvier, Ryan Amaudruz, Felix Arnold et al.

As AI workloads proliferate, optimizing arithmetic units is becoming increasingly important for reducing the footprint of digital systems. Conventional design flows, which often rely on manual or heuristic-based optimization, are limited in their ability to thoroughly explore the vast design space. In this paper, we introduce GENIAL, a machine learning-based framework for the automatic generation and optimization of arithmetic units, with a focus on multipliers. At the core of GENIAL is a Transformer-based surrogate model trained in two stages, involving self-supervised pretraining followed by supervised finetuning, to robustly forecast key hardware metrics such as power and area from abstracted design representations. By inverting the surrogate model, GENIAL efficiently searches for new operand encodings that directly minimize power consumption in arithmetic units for specific input data distributions. Extensive experiments on large datasets demonstrate that GENIAL is consistently more sample efficient than other methods, and converges faster towards optimized designs. This enables deployment of a high-effort logic synthesis optimization flow in the loop, improving the accuracy of the surrogate model. Notably, GENIAL automatically discovers encodings that achieve up to 18% switching activity savings within multipliers on representative AI workloads compared with the conventional two's complement. We also demonstrate the versatility of our approach by achieving significant improvements on Finite State Machines, highlighting GENIAL's applicability for a wide spectrum of logic functions. Together, these advances mark a significant step toward automated Quality-of-Results-optimized combinational circuit generation for digital systems.

Giuseppe Ruggeri, Renzo Andri, Daniele Jahier Pagliari et al.

Deep Recommender Models (DLRMs) inference is a fundamental AI workload accounting for more than 79% of the total AI workload in Meta's data centers. DLRMs' performance bottleneck is found in the embedding layers, which perform many random memory accesses to retrieve small embedding vectors from tables of various sizes. We propose the design of tailored data flows to speedup embedding look-ups. Namely, we propose four strategies to look up an embedding table effectively on one core, and a framework to automatically map the tables asymmetrically to the multiple cores of a SoC. We assess the effectiveness of our method using the Huawei Ascend AI accelerators, comparing it with the default Ascend compiler, and we perform high-level comparisons with Nvidia A100. Results show a speed-up varying from 1.5x up to 6.5x for real workload distributions, and more than 20x for extremely unbalanced distributions. Furthermore, the method proves to be much more independent of the query distribution than the baseline.

Lara Magdalena Lazier, Aritra Dhar, Vasilije Stambolic et al.

Corporate LLMs are gaining traction for efficient knowledge dissemination and management within organizations. However, as current LLMs are vulnerable to leaking sensitive information, it has proven difficult to apply them in settings where strict access control is necessary. To this end, we design AC-LoRA, an end-to-end system for access control-aware corporate LLM chatbots that maintains a strong information isolation guarantee. AC-LoRA maintains separate LoRA adapters for permissioned datasets, along with the document embedding they are finetuned on. AC-LoRA retrieves a precise set of LoRA adapters based on the similarity score with the user query and their permission. This similarity score is later used to merge the responses if more than one LoRA is retrieved, without requiring any additional training for LoRA routing. We provide an end-to-end prototype of AC-LoRA, evaluate it on two datasets, and show that AC-LoRA matches or even exceeds the performance of state-of-the-art LoRA mixing techniques while providing strong isolation guarantees. Furthermore, we show that AC-LoRA design can be directly applied to different modalities.

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.

Gianmarco Cerutti, Lukas Cavigelli, Renzo Andri et al.

Keyword spotting (KWS) is a crucial function enabling the interaction with the many ubiquitous smart devices in our surroundings, either activating them through wake-word or directly as a human-computer interface. For many applications, KWS is the entry point for our interactions with the device and, thus, an always-on workload. Many smart devices are mobile and their battery lifetime is heavily impacted by continuously running services. KWS and similar always-on services are thus the focus when optimizing the overall power consumption. This work addresses KWS energy-efficiency on low-cost microcontroller units (MCUs). We combine analog binary feature extraction with binary neural networks. By replacing the digital preprocessing with the proposed analog front-end, we show that the energy required for data acquisition and preprocessing can be reduced by 29x, cutting its share from a dominating 85% to a mere 16% of the overall energy consumption for our reference KWS application. Experimental evaluations on the Speech Commands Dataset show that the proposed system outperforms state-of-the-art accuracy and energy efficiency, respectively, by 1% and 4.3x on a 10-class dataset while providing a compelling accuracy-energy trade-off including a 2% accuracy drop for a 71x energy reduction.

Xavier Timoneda, Lukas Cavigelli

Logic optimization is an NP-hard problem commonly approached through hand-engineered heuristics. We propose to combine graph convolutional networks with reinforcement learning and a novel, scalable node embedding method to learn which local transforms should be applied to the logic graph. We show that this method achieves a similar size reduction as ABC on smaller circuits and outperforms it by 1.5-1.75x on larger random graphs.

Thorir Mar Ingolfsson, Xiaying Wang, Michael Hersche et al.

Personalized ubiquitous healthcare solutions require energy-efficient wearable platforms that provide an accurate classification of bio-signals while consuming low average power for long-term battery-operated use. Single lead electrocardiogram (ECG) signals provide the ability to detect, classify, and even predict cardiac arrhythmia. In this paper, we propose a novel temporal convolutional network (TCN) that achieves high accuracy while still being feasible for wearable platform use. Experimental results on the ECG5000 dataset show that the TCN has a similar accuracy (94.2%) score as the state-of-the-art (SoA) network while achieving an improvement of 16.5% in the balanced accuracy score. This accurate classification is done with 27 times fewer parameters and 37 times less multiply-accumulate operations. We test our implementation on two publicly available platforms, the STM32L475, which is based on ARM Cortex M4F, and the GreenWaves Technologies GAP8 on the GAPuino board, based on 1+8 RISC-V CV32E40P cores. Measurements show that the GAP8 implementation respects the real-time constraints while consuming 0.10 mJ per inference. With 9.91 GMAC/s/W, it is 23.0 times more energy-efficient and 46.85 times faster than an implementation on the ARM Cortex M4F (0.43 GMAC/s/W). Overall, we obtain 8.1% higher accuracy while consuming 19.6 times less energy and being 35.1 times faster compared to a previous SoA embedded implementation.

Gianmarco Cerutti, Renzo Andri, Lukas Cavigelli et al.

Sound event detection (SED) is a hot topic in consumer and smart city applications. Existing approaches based on Deep Neural Networks are very effective, but highly demanding in terms of memory, power, and throughput when targeting ultra-low power always-on devices. Latency, availability, cost, and privacy requirements are pushing recent IoT systems to process the data on the node, close to the sensor, with a very limited energy supply, and tight constraints on the memory size and processing capabilities precluding to run state-of-the-art DNNs. In this paper, we explore the combination of extreme quantization to a small-footprint binary neural network (BNN) with the highly energy-efficient, RISC-V-based (8+1)-core GAP8 microcontroller. Starting from an existing CNN for SED whose footprint (815 kB) exceeds the 512 kB of memory available on our platform, we retrain the network using binary filters and activations to match these memory constraints. (Fully) binary neural networks come with a natural drop in accuracy of 12-18% on the challenging ImageNet object recognition challenge compared to their equivalent full-precision baselines. This BNN reaches a 77.9% accuracy, just 7% lower than the full-precision version, with 58 kB (7.2 times less) for the weights and 262 kB (2.4 times less) memory in total. With our BNN implementation, we reach a peak throughput of 4.6 GMAC/s and 1.5 GMAC/s over the full network, including preprocessing with Mel bins, which corresponds to an efficiency of 67.1 GMAC/s/W and 31.3 GMAC/s/W, respectively. Compared to the performance of an ARM Cortex-M4 implementation, our system has a 10.3 times faster execution time and a 51.1 times higher energy-efficiency.

Thorir Mar Ingolfsson, Michael Hersche, Xiaying Wang et al.

In recent years, deep learning (DL) has contributed significantly to the improvement of motor-imagery brain-machine interfaces (MI-BMIs) based on electroencephalography(EEG). While achieving high classification accuracy, DL models have also grown in size, requiring a vast amount of memory and computational resources. This poses a major challenge to an embedded BMI solution that guarantees user privacy, reduced latency, and low power consumption by processing the data locally. In this paper, we propose EEG-TCNet, a novel temporal convolutional network (TCN) that achieves outstanding accuracy while requiring few trainable parameters. Its low memory footprint and low computational complexity for inference make it suitable for embedded classification on resource-limited devices at the edge. Experimental results on the BCI Competition IV-2a dataset show that EEG-TCNet achieves 77.35% classification accuracy in 4-class MI. By finding the optimal network hyperparameters per subject, we further improve the accuracy to 83.84%. Finally, we demonstrate the versatility of EEG-TCNet on the Mother of All BCI Benchmarks (MOABB), a large scale test benchmark containing 12 different EEG datasets with MI experiments. The results indicate that EEG-TCNet successfully generalizes beyond one single dataset, outperforming the current state-of-the-art (SoA) on MOABB by a meta-effect of 0.25.

Tibor Schneider, Xiaying Wang, Michael Hersche et al.

Motor-Imagery Brain--Machine Interfaces (MI-BMIs)promise direct and accessible communication between human brains and machines by analyzing brain activities recorded with Electroencephalography (EEG). Latency, reliability, and privacy constraints make it unsuitable to offload the computation to the cloud. Practical use cases demand a wearable, battery-operated device with low average power consumption for long-term use. Recently, sophisticated algorithms, in particular deep learning models, have emerged for classifying EEG signals. While reaching outstanding accuracy, these models often exceed the limitations of edge devices due to their memory and computational requirements. In this paper, we demonstrate algorithmic and implementation optimizations for EEGNET, a compact Convolutional Neural Network (CNN) suitable for many BMI paradigms. We quantize weights and activations to 8-bit fixed-point with a negligible accuracy loss of 0.4% on 4-class MI, and present an energy-efficient hardware-aware implementation on the Mr.Wolf parallel ultra-low power (PULP) System-on-Chip (SoC) by utilizing its custom RISC-V ISA extensions and 8-core compute cluster. With our proposed optimization steps, we can obtain an overall speedup of 64x and a reduction of up to 85% in memory footprint with respect to a single-core layer-wise baseline implementation. Our implementation takes only 5.82 ms and consumes 0.627 mJ per inference. With 21.0GMAC/s/W, it is 256x more energy-efficient than an EEGNET implementation on an ARM Cortex-M7 (0.082GMAC/s/W).

Lukas Cavigelli, Luca Benini

We present Random Partition Relaxation (RPR), a method for strong quantization of neural networks weight to binary (+1/-1) and ternary (+1/0/-1) values. Starting from a pre-trained model, we quantize the weights and then relax random partitions of them to their continuous values for retraining before re-quantizing them and switching to another weight partition for further adaptation. We demonstrate binary and ternary-weight networks with accuracies beyond the state-of-the-art for GoogLeNet and competitive performance for ResNet-18 and ResNet-50 using an SGD-based training method that can easily be integrated into existing frameworks.

Xiaying Wang, Lukas Cavigelli, Manuel Eggimann et al.

Synthetic aperture radar (SAR) data is becoming increasingly available to a wide range of users through commercial service providers with resolutions reaching 0.5m/px. Segmenting SAR data still requires skilled personnel, limiting the potential for large-scale use. We show that it is possible to automatically and reliably perform urban scene segmentation from next-gen resolution SAR data (0.15m/px) using deep neural networks (DNNs), achieving a pixel accuracy of 95.19% and a mean IoU of 74.67% with data collected over a region of merely 2.2km${}^2$. The presented DNN is not only effective, but is very small with only 63k parameters and computationally simple enough to achieve a throughput of around 500Mpx/s using a single GPU. We further identify that additional SAR receive antennas and data from multiple flights massively improve the segmentation accuracy. We describe a procedure for generating a high-quality segmentation ground truth from multiple inaccurate building and road annotations, which has been crucial to achieving these segmentation results.

Lukas Cavigelli, Georg Rutishauser, Luca Benini

In the wake of the success of convolutional neural networks in image classification, object recognition, speech recognition, etc., the demand for deploying these compute-intensive ML models on embedded and mobile systems with tight power and energy constraints at low cost, as well as for boosting throughput in data centers, is growing rapidly. This has sparked a surge of research into specialized hardware accelerators. Their performance is typically limited by I/O bandwidth, power consumption is dominated by I/O transfers to off-chip memory, and on-chip memories occupy a large part of the silicon area. We introduce and evaluate a novel, hardware-friendly, and lossless compression scheme for the feature maps present within convolutional neural networks. We present hardware architectures and synthesis results for the compressor and decompressor in 65nm. With a throughput of one 8-bit word/cycle at 600MHz, they fit into 2.8kGE and 3.0kGE of silicon area, respectively - together the size of less than seven 8-bit multiply-add units at the same throughput. We show that an average compression ratio of 5.1x for AlexNet, 4x for VGG-16, 2.4x for ResNet-34 and 2.2x for MobileNetV2 can be achieved - a gain of 45-70% over existing methods. Our approach also works effectively for various number formats, has a low frame-to-frame variance on the compression ratio, and achieves compression factors for gradient map compression during training that are even better than for inference.

Matteo Spallanzani, Lukas Cavigelli, Gian Paolo Leonardi et al.

We present a theoretical and experimental investigation of the quantization problem for artificial neural networks. We provide a mathematical definition of quantized neural networks and analyze their approximation capabilities, showing in particular that any Lipschitz-continuous map defined on a hypercube can be uniformly approximated by a quantized neural network. We then focus on the regularization effect of additive noise on the arguments of multi-step functions inherent to the quantization of continuous variables. In particular, when the expectation operator is applied to a non-differentiable multi-step random function, and if the underlying probability density is differentiable (in either classical or weak sense), then a differentiable function is retrieved, with explicit bounds on its Lipschitz constant. Based on these results, we propose a novel gradient-based training algorithm for quantized neural networks that generalizes the straight-through estimator, acting on noise applied to the network's parameters. We evaluate our algorithm on the CIFAR-10 and ImageNet image classification benchmarks, showing state-of-the-art performance on AlexNet and MobileNetV2 for ternary networks.

Lukas Cavigelli, Luca Benini

After the tremendous success of convolutional neural networks in image classification, object detection, speech recognition, etc., there is now rising demand for deployment of these compute-intensive ML models on tightly power constrained embedded and mobile systems at low cost as well as for pushing the throughput in data centers. This has triggered a wave of research towards specialized hardware accelerators. Their performance is often constrained by I/O bandwidth and the energy consumption is dominated by I/O transfers to off-chip memory. We introduce and evaluate a novel, hardware-friendly compression scheme for the feature maps present within convolutional neural networks. We show that an average compression ratio of 4.4x relative to uncompressed data and a gain of 60% over existing method can be achieved for ResNet-34 with a compression block requiring <300 bit of sequential cells and minimal combinational logic.

Lukas Cavigelli, Luca Benini

The last few years have brought advances in computer vision at an amazing pace, grounded on new findings in deep neural network construction and training as well as the availability of large labeled datasets. Applying these networks to images demands a high computational effort and pushes the use of state-of-the-art networks on real-time video data out of reach of embedded platforms. Many recent works focus on reducing network complexity for real-time inference on embedded computing platforms. We adopt an orthogonal viewpoint and propose a novel algorithm exploiting the spatio-temporal sparsity of pixel changes. This optimized inference procedure resulted in an average speed-up of 9.1x over cuDNN on the Tegra X2 platform at a negligible accuracy loss of <0.1% and no retraining of the network for a semantic segmentation application. Similarly, an average speed-up of 7.0x has been achieved for a pose detection DNN and a reduction of 5x of the number of arithmetic operations to be performed for object detection on static camera video surveillance data. These throughput gains combined with a lower power consumption result in an energy efficiency of 511 GOp/s/W compared to 70 GOp/s/W for the baseline.

Renzo Andri, Lukas Cavigelli, Davide Rossi et al.

Deep neural networks have achieved impressive results in computer vision and machine learning. Unfortunately, state-of-the-art networks are extremely compute and memory intensive which makes them unsuitable for mW-devices such as IoT end-nodes. Aggressive quantization of these networks dramatically reduces the computation and memory footprint. Binary-weight neural networks (BWNs) follow this trend, pushing weight quantization to the limit. Hardware accelerators for BWNs presented up to now have focused on core efficiency, disregarding I/O bandwidth and system-level efficiency that are crucial for deployment of accelerators in ultra-low power devices. We present Hyperdrive: a BWN accelerator dramatically reducing the I/O bandwidth exploiting a novel binary-weight streaming approach, which can be used for arbitrarily sized convolutional neural network architecture and input resolution by exploiting the natural scalability of the compute units both at chip-level and system-level by arranging Hyperdrive chips systolically in a 2D mesh while processing the entire feature map together in parallel. Hyperdrive achieves 4.3 TOp/s/W system-level efficiency (i.e., including I/Os)---3.1x higher than state-of-the-art BWN accelerators, even if its core uses resource-intensive FP16 arithmetic for increased robustness.

Andrawes Al Bahou, Geethan Karunaratne, Renzo Andri et al.

Deploying state-of-the-art CNNs requires power-hungry processors and off-chip memory. This precludes the implementation of CNNs in low-power embedded systems. Recent research shows CNNs sustain extreme quantization, binarizing their weights and intermediate feature maps, thereby saving 8-32\x memory and collapsing energy-intensive sum-of-products into XNOR-and-popcount operations. We present XNORBIN, an accelerator for binary CNNs with computation tightly coupled to memory for aggressive data reuse. Implemented in UMC 65nm technology XNORBIN achieves an energy efficiency of 95 TOp/s/W and an area efficiency of 2.0 TOp/s/MGE at 0.8 V.

Manuele Rusci, Lukas Cavigelli, Luca Benini

Design automation in general, and in particular logic synthesis, can play a key role in enabling the design of application-specific Binarized Neural Networks (BNN). This paper presents the hardware design and synthesis of a purely combinational BNN for ultra-low power near-sensor processing. We leverage the major opportunities raised by BNN models, which consist mostly of logical bit-wise operations and integer counting and comparisons, for pushing ultra-low power deep learning circuits close to the sensor and coupling it with binarized mixed-signal image sensor data. We analyze area, power and energy metrics of BNNs synthesized as combinational networks. Our synthesis results in GlobalFoundries 22nm SOI technology shows a silicon area of 2.61mm2 for implementing a combinational BNN with 32x32 binary input sensor receptive field and weight parameters fixed at design time. This is 2.2x smaller than a synthesized network with re-configurable parameters. With respect to other comparable techniques for deep learning near-sensor processing, our approach features a 10x higher energy efficiency.

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.

Matthias Meyer, Lukas Cavigelli, Lothar Thiele

Wireless distributed systems as used in sensor networks, Internet-of-Things and cyber-physical systems, impose high requirements on resource efficiency. Advanced preprocessing and classification of data at the network edge can help to decrease the communication demand and to reduce the amount of data to be processed centrally. In the area of distributed acoustic sensing, the combination of algorithms with a high classification rate and resource-constraint embedded systems is essential. Unfortunately, algorithms for acoustic event detection have a high memory and computational demand and are not suited for execution at the network edge. This paper addresses these aspects by applying structural optimizations to a convolutional neural network for audio event detection to reduce the memory requirement by a factor of more than 500 and the computational effort by a factor of 2.1 while performing 9.2% better.

Lukas Cavigelli, Philippe Degen, Luca Benini

Extracting per-frame features using convolutional neural networks for real-time processing of video data is currently mainly performed on powerful GPU-accelerated workstations and compute clusters. However, there are many applications such as smart surveillance cameras that require or would benefit from on-site processing. To this end, we propose and evaluate a novel algorithm for change-based evaluation of CNNs for video data recorded with a static camera setting, exploiting the spatio-temporal sparsity of pixel changes. We achieve an average speed-up of 8.6x over a cuDNN baseline on a realistic benchmark with a negligible accuracy loss of less than 0.1% and no retraining of the network. The resulting energy efficiency is 10x higher than that of per-frame evaluation and reaches an equivalent of 328 GOp/s/W on the Tegra X1 platform.

Eirikur Agustsson, Fabian Mentzer, Michael Tschannen et al.

We present a new approach to learn compressible representations in deep architectures with an end-to-end training strategy. Our method is based on a soft (continuous) relaxation of quantization and entropy, which we anneal to their discrete counterparts throughout training. We showcase this method for two challenging applications: Image compression and neural network compression. While these tasks have typically been approached with different methods, our soft-to-hard quantization approach gives results competitive with the state-of-the-art for both.

Lukas Cavigelli, Pascal Hager, Luca Benini

Lossy image compression algorithms are pervasively used to reduce the size of images transmitted over the web and recorded on data storage media. However, we pay for their high compression rate with visual artifacts degrading the user experience. Deep convolutional neural networks have become a widespread tool to address high-level computer vision tasks very successfully. Recently, they have found their way into the areas of low-level computer vision and image processing to solve regression problems mostly with relatively shallow networks. We present a novel 12-layer deep convolutional network for image compression artifact suppression with hierarchical skip connections and a multi-scale loss function. We achieve a boost of up to 1.79 dB in PSNR over ordinary JPEG and an improvement of up to 0.36 dB over the best previous ConvNet result. We show that a network trained for a specific quality factor (QF) is resilient to the QF used to compress the input image - a single network trained for QF 60 provides a PSNR gain of more than 1.5 dB over the wide QF range from 40 to 76.

Lukas Cavigelli, Dominic Bernath, Michele Magno et al.

Detecting and classifying targets in video streams from surveillance cameras is a cumbersome, error-prone and expensive task. Often, the incurred costs are prohibitive for real-time monitoring. This leads to data being stored locally or transmitted to a central storage site for post-incident examination. The required communication links and archiving of the video data are still expensive and this setup excludes preemptive actions to respond to imminent threats. An effective way to overcome these limitations is to build a smart camera that transmits alerts when relevant video sequences are detected. Deep neural networks (DNNs) have come to outperform humans in visual classifications tasks. The concept of DNNs and Convolutional Networks (ConvNets) can easily be extended to make use of higher-dimensional input data such as multispectral data. We explore this opportunity in terms of achievable accuracy and required computational effort. To analyze the precision of DNNs for scene labeling in an urban surveillance scenario we have created a dataset with 8 classes obtained in a field experiment. We combine an RGB camera with a 25-channel VIS-NIR snapshot sensor to assess the potential of multispectral image data for target classification. We evaluate several new DNNs, showing that the spectral information fused together with the RGB frames can be used to improve the accuracy of the system or to achieve similar accuracy with a 3x smaller computation effort. We achieve a very high per-pixel accuracy of 99.1%. Even for scarcely occurring, but particularly interesting classes, such as cars, 75% of the pixels are labeled correctly with errors occurring only around the border of the objects. This high accuracy was obtained with a training set of only 30 labeled images, paving the way for fast adaptation to various application scenarios.

Michael Tschannen, Lukas Cavigelli, Fabian Mentzer et al.

We propose a highly structured neural network architecture for semantic segmentation with an extremely small model size, suitable for low-power embedded and mobile platforms. Specifically, our architecture combines i) a Haar wavelet-based tree-like convolutional neural network (CNN), ii) a random layer realizing a radial basis function kernel approximation, and iii) a linear classifier. While stages i) and ii) are completely pre-specified, only the linear classifier is learned from data. We apply the proposed architecture to outdoor scene and aerial image semantic segmentation and show that the accuracy of our architecture is competitive with conventional pixel classification CNNs. Furthermore, we demonstrate that the proposed architecture is data efficient in the sense of matching the accuracy of pixel classification CNNs when trained on a much smaller data set.

Renzo Andri, Lukas Cavigelli, Davide Rossi et al.

Convolutional neural networks (CNNs) have revolutionized the world of computer vision over the last few years, pushing image classification beyond human accuracy. The computational effort of today's CNNs requires power-hungry parallel processors or GP-GPUs. Recent developments in CNN accelerators for system-on-chip integration have reduced energy consumption significantly. Unfortunately, even these highly optimized devices are above the power envelope imposed by mobile and deeply embedded applications and face hard limitations caused by CNN weight I/O and storage. This prevents the adoption of CNNs in future ultra-low power Internet of Things end-nodes for near-sensor analytics. Recent algorithmic and theoretical advancements enable competitive classification accuracy even when limiting CNNs to binary (+1/-1) weights during training. These new findings bring major optimization opportunities in the arithmetic core by removing the need for expensive multiplications, as well as reducing I/O bandwidth and storage. In this work, we present an accelerator optimized for binary-weight CNNs that achieves 1510 GOp/s at 1.2 V on a core area of only 1.33 MGE (Million Gate Equivalent) or 0.19 mm$^2$ and with a power dissipation of 895 μW in UMC 65 nm technology at 0.6 V. Our accelerator significantly outperforms the state-of-the-art in terms of energy and area efficiency achieving 61.2 TOp/s/W@0.6 V and 1135 GOp/s/MGE@1.2 V, respectively.

Lukas Cavigelli, Luca Benini

An ever increasing number of computer vision and image/video processing challenges are being approached using deep convolutional neural networks, obtaining state-of-the-art results in object recognition and detection, semantic segmentation, action recognition, optical flow and superresolution. Hardware acceleration of these algorithms is essential to adopt these improvements in embedded and mobile computer vision systems. We present a new architecture, design and implementation as well as the first reported silicon measurements of such an accelerator, outperforming previous work in terms of power-, area- and I/O-efficiency. The manufactured device provides up to 196 GOp/s on 3.09 mm^2 of silicon in UMC 65nm technology and can achieve a power efficiency of 803 GOp/s/W. The massively reduced bandwidth requirements make it the first architecture scalable to TOp/s performance.