VEXP: A Low-Cost RISC-V ISA Extension for Accelerated Softmax Computation in Transformers

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.

Enhancing Autonomous Driving Systems with On-Board Deployed Large Language Models

Neural Networks (NNs) trained through supervised learning struggle with managing edge-case scenarios common in real-world driving due to the intractability of exhaustive datasets covering all edge-cases, making knowledge-driven approaches, akin to how humans intuitively detect unexpected driving behavior, a suitable complement to data-driven methods. This work proposes a hybrid architecture combining low-level Model Predictive Controller (MPC) with locally deployed Large Language Models (LLMs) to enhance decision-making and Human Machine Interaction (HMI). The DecisionxLLM module evaluates robotic state information against natural language instructions to ensure adherence to desired driving behavior. The MPCxLLM module then adjusts MPC parameters based on LLM-generated insights, achieving control adaptability while preserving the safety and constraint guarantees of traditional MPC systems. Further, to enable efficient on-board deployment and to eliminate dependency on cloud connectivity, we shift processing to the on-board computing platform: We propose an approach that exploits Retrieval Augmented Generation (RAG), Low Rank Adaptation (LoRA) fine-tuning, and quantization. Experimental results demonstrate that these enhancements yield significant improvements in reasoning accuracy by up to 10.45%, control adaptability by as much as 52.2%, and up to 10.5x increase in computational efficiency (tokens/s), validating the proposed framework's practicality for real-time deployment even on down-scaled robotic platforms. This work bridges high-level decision-making with low-level control adaptability, offering a synergistic framework for knowledge-driven and adaptive Autonomous Driving Systems (ADS).

FastVAR: Linear Visual Autoregressive Modeling via Cached Token Pruning

Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by 2.7$\times$ with negligible performance drop of <1%. We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one 2K image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR.

Q-MambaIR: Accurate Quantized Mamba for Efficient Image Restoration

State-Space Models (SSMs) have attracted considerable attention in Image Restoration (IR) due to their ability to scale linearly sequence length while effectively capturing long-distance dependencies. However, deploying SSMs to edge devices is challenging due to the constraints in memory, computing capacity, and power consumption, underscoring the need for efficient compression strategies. While low-bit quantization is an efficient model compression strategy for reducing size and accelerating IR tasks, SSM suffers substantial performance drops at ultra-low bit-widths (2-4 bits), primarily due to outliers that exacerbate quantization error. To address this challenge, we propose Q-MambaIR, an accurate, efficient, and flexible Quantized Mamba for IR tasks. Specifically, we introduce a Statistical Dynamic-balancing Learnable Scalar (DLS) to dynamically adjust the quantization mapping range, thereby mitigating the peak truncation loss caused by extreme values. Furthermore, we design a Range-floating Flexible Allocator (RFA) with an adaptive threshold to flexibly round values. This approach preserves high-frequency details and maintains the SSM's feature extraction capability. Notably, RFA also enables pre-deployment weight quantization, striking a balance between computational efficiency and model accuracy. Extensive experiments on IR tasks demonstrate that Q-MambaIR consistently outperforms existing quantized SSMs, achieving much higher state-of-the-art (SOTA) accuracy results with only a negligible increase in training computation and storage saving.

CamSAM2: Segment Anything Accurately in Camouflaged Videos

Video camouflaged object segmentation (VCOS), aiming at segmenting camouflaged objects that seamlessly blend into their environment, is a fundamental vision task with various real-world applications. With the release of SAM2, video segmentation has witnessed significant progress. However, SAM2's capability of segmenting camouflaged videos is suboptimal, especially when given simple prompts such as point and box. To address the problem, we propose Camouflaged SAM2 (CamSAM2), which enhances SAM2's ability to handle camouflaged scenes without modifying SAM2's parameters. Specifically, we introduce a decamouflaged token to provide the flexibility of feature adjustment for VCOS. To make full use of fine-grained and high-resolution features from the current frame and previous frames, we propose implicit object-aware fusion (IOF) and explicit object-aware fusion (EOF) modules, respectively. Object prototype generation (OPG) is introduced to abstract and memorize object prototypes with informative details using high-quality features from previous frames. Extensive experiments are conducted to validate the effectiveness of our approach. While CamSAM2 only adds negligible learnable parameters to SAM2, it substantially outperforms SAM2 on three VCOS datasets, especially achieving 12.2 mDice gains with click prompt on MoCA-Mask and 19.6 mDice gains with mask prompt on SUN-SEG-Hard, with Hiera-T as the backbone. The code will be available at https://github.com/zhoustan/CamSAM2.

Towards Extreme Pruning of LLMs with Plug-and-Play Mixed Sparsity

N:M structured pruning is essential for large language models (LLMs) because it can remove less important network weights and reduce the memory and computation requirements. Existing pruning methods mainly focus on designing metrics to measure the importance of network components to guide pruning. Apart from the impact of these metrics, we observe that different layers have different sensitivities over the network performance. Thus, we propose an efficient method based on the trace of Fisher Information Matrix (FIM) to quantitatively measure and verify the different sensitivities across layers. Based on this, we propose Mixed Sparsity Pruning (MSP) which uses a pruning-oriented evolutionary algorithm (EA) to determine the optimal sparsity levels for different layers. To guarantee fast convergence and achieve promising performance, we utilize efficient FIM-inspired layer-wise sensitivity to initialize the population of EA. In addition, our MSP can work as a plug-and-play module, ready to be integrated into existing pruning methods. Extensive experiments on LLaMA and LLaMA-2 on language modeling and zero-shot tasks demonstrate our superior performance. In particular, in extreme pruning ratio (e.g. 75%), our method significantly outperforms existing methods in terms of perplexity (PPL) by orders of magnitude (Figure 1).

Lightweight Software Kernels and Hardware Extensions for Efficient Sparse Deep Neural Networks on Microcontrollers

The acceleration of pruned Deep Neural Networks (DNNs) on edge devices such as Microcontrollers (MCUs) is a challenging task, given the tight area- and power-constraints of these devices. In this work, we propose a three-fold contribution to address this problem. First, we design a set of optimized software kernels for N:M pruned layers, targeting ultra-low-power, multicore RISC-V MCUs, which are up to 2.1x and 3.4x faster than their dense counterparts at 1:8 and 1:16 sparsity, respectively. Then, we implement a lightweight Instruction-Set Architecture (ISA) extension to accelerate the indirect load and non-zero indices decompression operations required by our kernels, obtaining up to 1.9x extra speedup, at the cost of a 5% area overhead. Lastly, we extend an open-source DNN compiler to utilize our sparse kernels for complete networks, showing speedups of 3.21x and 1.81x on a ResNet18 and a Vision Transformer (ViT), with less than 1.5% accuracy drop compared to a dense baseline.

FEMBA: Efficient and Scalable EEG Analysis with a Bidirectional Mamba Foundation Model

Accurate and efficient electroencephalography (EEG) analysis is essential for detecting seizures and artifacts in long-term monitoring, with applications spanning hospital diagnostics to wearable health devices. Robust EEG analytics have the potential to greatly improve patient care. However, traditional deep learning models, especially Transformer-based architectures, are hindered by their quadratic time and memory complexity, making them less suitable for resource-constrained environments. To address these challenges, we present FEMBA (Foundational EEG Mamba + Bidirectional Architecture), a novel self-supervised framework that establishes new efficiency benchmarks for EEG analysis through bidirectional state-space modeling. Unlike Transformer-based models, which incur quadratic time and memory complexity, FEMBA scales linearly with sequence length, enabling more scalable and efficient processing of extended EEG recordings. Trained on over 21,000 hours of unlabeled EEG and fine-tuned on three downstream tasks, FEMBA achieves competitive performance in comparison with transformer models, with significantly lower computational cost. Specifically, it reaches 81.82% balanced accuracy (0.8921 AUROC) on TUAB and 0.949 AUROC on TUAR, while a tiny 7.8M-parameter variant demonstrates viability for resource-constrained devices. These results pave the way for scalable, general-purpose EEG analytics in both clinical and highlight FEMBA as a promising candidate for wearable applications.

CEReBrO: Compact Encoder for Representations of Brain Oscillations Using Efficient Alternating Attention

Electroencephalograph (EEG) is a crucial tool for studying brain activity. Recently, self-supervised learning methods leveraging large unlabeled datasets have emerged as a potential solution to the scarcity of widely available annotated EEG data. However, current methods suffer from at least one of the following limitations: i) sub-optimal EEG signal modeling, ii) model sizes in the hundreds of millions of trainable parameters, and iii) reliance on private datasets and/or inconsistent public benchmarks, hindering reproducibility. To address these challenges, we introduce a Compact Encoder for Representations of Brain Oscillations using alternating attention (CEReBrO), a new small EEG foundation model. Our tokenization scheme represents EEG signals at a per-channel patch granularity. We propose an alternating attention mechanism that jointly models intra-channel temporal dynamics and inter-channel spatial correlations, achieving 2x speed improvement with 6x less memory required compared to standard self-attention. We present several model sizes ranging from 3.6 million to 85 million parameters. Pre-trained on over 20,000 hours of publicly available scalp EEG recordings with diverse channel configurations, our models set new benchmarks in emotion detection and seizure detection tasks, with competitive performance in anomaly classification and gait prediction. This validates our models' effectiveness and effictiveness.

Open-Source Heterogeneous SoCs for AI: The PULP Platform Experience

Since 2013, the PULP (Parallel Ultra-Low Power) Platform project has been one of the most active and successful initiatives in designing research IPs and releasing them as open-source. Its portfolio now ranges from processor cores to network-on-chips, peripherals, SoC templates, and full hardware accelerators. In this article, we focus on the PULP experience designing heterogeneous AI acceleration SoCs - an endeavour encompassing SoC architecture definition; development, verification, and integration of acceleration IPs; front- and back-end VLSI design; testing; development of AI deployment software.