Wayne Luk

Zhiqiang Que, Hongxiang Fan, Marcus Loo et al.

This work presents a novel reconfigurable architecture for Low Latency Graph Neural Network (LL-GNN) designs for particle detectors, delivering unprecedented low latency performance. Incorporating FPGA-based GNNs into particle detectors presents a unique challenge since it requires sub-microsecond latency to deploy the networks for online event selection with a data rate of hundreds of terabytes per second in the Level-1 triggers at the CERN Large Hadron Collider experiments. This paper proposes a novel outer-product based matrix multiplication approach, which is enhanced by exploiting the structured adjacency matrix and a column-major data layout. Moreover, a fusion step is introduced to further reduce the end-to-end design latency by eliminating unnecessary boundaries. Furthermore, a GNN-specific algorithm-hardware co-design approach is presented which not only finds a design with a much better latency but also finds a high accuracy design under given latency constraints. To facilitate this, a customizable template for this low latency GNN hardware architecture has been designed and open-sourced, which enables the generation of low-latency FPGA designs with efficient resource utilization using a high-level synthesis tool. Evaluation results show that our FPGA implementation is up to 9.0 times faster and achieves up to 13.1 times higher power efficiency than a GPU implementation. Compared to the previous FPGA implementations, this work achieves 6.51 to 16.7 times lower latency. Moreover, the latency of our FPGA design is sufficiently low to enable deployment of GNNs in a sub-microsecond, real-time collider trigger system, enabling it to benefit from improved accuracy. The proposed LL-GNN design advances the next generation of trigger systems by enabling sophisticated algorithms to process experimental data efficiently.

Chang Sun, Zhiqiang Que, Bakhtiar Zadeh et al.

Lookup-table (LUT) based neural networks can deliver ultra-low latency and excellent hardware efficiency on FPGAs by mapping arithmetic operations directly onto the logic primitives. However, state-of-the-art LUT-aware training (LAT) approaches remain difficult to use in practice: they are often orders of magnitude slower to train than conventional networks, require non-trivial manual tuning for hardware efficiency, and lack an end-to-end workflow. This work presents HGQ-LUT, integrated in https://github.com/calad0i/HGQ2, a new LAT approach that achieves state-of-the-art hardware efficiency while accelerating training by over 100 times on modern GPUs. HGQ-LUT introduces LUT-Dense and LUT-Conv layers that are implemented with regular, accelerator-efficient tensor operations during training, which are then compiled into logic LUTs for hardware. By combining these layers with fine-grained, element-wise heterogeneous quantization (including zero-bit pruning) and a LUT-aware resource surrogate, HGQ-LUT enables the automatic exploration of accuracy-resource trade-offs without manual bit-width tuning. We further integrate HGQ-LUT into open-source toolchains, enabling unified design, compilation, and bit-exact verification of hybrid architectures that mix LUT-based with conventional arithmetic blocks. These features make LAT-based DNNs practical for real-world deployment, such as at the CERN Large Hadron Collider's experiments.

Ruoqing Zheng, Chang Sun, Qibin Liu et al.

We present JetFormer, a versatile and scalable encoder-only Transformer architecture for particle jet tagging at the Large Hadron Collider (LHC). Unlike prior approaches that are often tailored to specific deployment regimes, JetFormer is designed to operate effectively across the full spectrum of jet tagging scenarios, from high-accuracy offline analysis to ultra-low-latency online triggering. The model processes variable-length sets of particle features without relying on input of explicit pairwise interactions, yet achieves competitive or superior performance compared to state-of-the-art methods. On the large-scale JetClass dataset, a large-scale JetFormer matches the accuracy of the interaction-rich ParT model (within 0.7%) while using 37.4% fewer FLOPs, demonstrating its computational efficiency and strong generalization. On benchmark HLS4ML 150P datasets, JetFormer consistently outperforms existing models such as MLPs, Deep Sets, and Interaction Networks by 3-4% in accuracy. To bridge the gap to hardware deployment, we further introduce a hardware-aware optimization pipeline based on multi-objective hyperparameter search, yielding compact variants like JetFormer-tiny suitable for FPGA-based trigger systems with sub-microsecond latency requirements. Through structured pruning and quantization, we show that JetFormer can be aggressively compressed with minimal accuracy loss. By unifying high-performance modeling and deployability within a single architectural framework, JetFormer provides a practical pathway for deploying Transformer-based jet taggers in both offline and online environments at the LHC. Code is available at https://github.com/walkieq/JetFormer.

Jan-Frederik Schulte, Benjamin Ramhorst, Chang Sun et al.

We present hls4ml, a free and open-source platform that translates machine learning (ML) models from modern deep learning frameworks into high-level synthesis (HLS) code that can be integrated into full designs for field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs). With its flexible and modular design, hls4ml supports a large number of deep learning frameworks and can target HLS compilers from several vendors, including Vitis HLS, Intel oneAPI and Catapult HLS. Together with a wider eco-system for software-hardware co-design, hls4ml has enabled the acceleration of ML inference in a wide range of commercial and scientific applications where low latency, resource usage, and power consumption are critical. In this paper, we describe the structure and functionality of the hls4ml platform. The overarching design considerations for the generated HLS code are discussed, together with selected performance results.

Ebby Samson, Naveen Mellempudi, Wayne Luk et al.

A number of companies recently worked together to release the new Open Compute Project MX standard for low-precision computation, aimed at efficient neural network implementation. In this paper, we describe and evaluate the first open-source FPGA implementation of the arithmetic defined in the standard. Our designs fully support all the standard's concrete formats for conversion into and out of MX formats and for the standard-defined arithmetic operations, as well as arbitrary fixed-point and floating-point formats. Certain elements of the standard are left as implementation-defined, and we present the first concrete FPGA-inspired choices for these elements, which we outline in the paper. Our library of optimized hardware components is available open source, and can be used to build larger systems. For this purpose, we also describe and release an open-source Pytorch library for quantization into the new standard, integrated with the Brevitas library so that the community can develop novel neural network designs quantized with MX formats in mind. We demonstrate the usability and efficacy of our libraries via the implementation of example neural networks such as ResNet-18 on the ImageNet ILSVRC12 dataset. Our testing shows that MX is very effective for formats such as INT5 or FP6 which are not natively supported on GPUs. This gives FPGAs an advantage as they have the flexibility to implement a custom datapath and take advantage of the smaller area footprints offered by these formats.

Haoran Wu, Can Xiao, Jiayi Nie et al.

LLMs now form the backbone of AI agents across a diverse range of applications, including tool use, command-line interfaces, and web or computer interaction. These agentic LLM inference tasks are fundamentally different from chatbot-focused inference. They often involve much longer context lengths to capture complex and prolonged inputs, such as an entire webpage DOM or complicated tool-call trajectories. This, in turn, generates significant off-chip memory traffic during inference and causes workloads to be constrained by two memory walls, namely the bandwidth wall and the capacity wall, preventing compute units from achieving high utilization. In this paper, we introduce PLENA, a hardware-software co-designed system built around three core optimization pathways. PLENA features a novel flattened systolic-array architecture (Pathway 1) and efficient compute and memory units that support an asymmetric quantization scheme (Pathway 2). It also provides native support for FlashAttention (Pathway 3). In addition, PLENA includes a complete software-hardware stack, consisting of a custom ISA, a compiler, a transaction-level simulator, and an automated design-space exploration flow. Experimental results show that PLENA delivers up to 2.23x and 4.70x higher throughput than the A100 GPU and TPU v6e, respectively, under identical multiplier counts and memory configurations during LLaMA agentic inference. PLENA also achieves up to 4.04x higher energy efficiency than the A100 GPU. The full PLENA system, including its simulator, compiler, ISA, and RTL implementation, will be open-sourced to the research community.

Hongxiang Fan, Thomas Chau, Stylianos I. Venieris et al.

Attention-based neural networks have become pervasive in many AI tasks. Despite their excellent algorithmic performance, the use of the attention mechanism and feed-forward network (FFN) demands excessive computational and memory resources, which often compromises their hardware performance. Although various sparse variants have been introduced, most approaches only focus on mitigating the quadratic scaling of attention on the algorithm level, without explicitly considering the efficiency of mapping their methods on real hardware designs. Furthermore, most efforts only focus on either the attention mechanism or the FFNs but without jointly optimizing both parts, causing most of the current designs to lack scalability when dealing with different input lengths. This paper systematically considers the sparsity patterns in different variants from a hardware perspective. On the algorithmic level, we propose FABNet, a hardware-friendly variant that adopts a unified butterfly sparsity pattern to approximate both the attention mechanism and the FFNs. On the hardware level, a novel adaptable butterfly accelerator is proposed that can be configured at runtime via dedicated hardware control to accelerate different butterfly layers using a single unified hardware engine. On the Long-Range-Arena dataset, FABNet achieves the same accuracy as the vanilla Transformer while reducing the amount of computation by 10 to 66 times and the number of parameters 2 to 22 times. By jointly optimizing the algorithm and hardware, our FPGA-based butterfly accelerator achieves 14.2 to 23.2 times speedup over state-of-the-art accelerators normalized to the same computational budget. Compared with optimized CPU and GPU designs on Raspberry Pi 4 and Jetson Nano, our system is up to 273.8 and 15.1 times faster under the same power budget.

Tong Zhao, Ce Guo, Wayne Luk et al.

Causal discovery with instantaneous effects in multivariate time series is challenging, as the instantaneous structure must be acyclic. Prior methods enforce this by either separating instantaneous and lagged estimation into multi-stage pipelines or imposing algebraic acyclicity constraints via complex augmented Lagrangian optimization, both of which incur high computational cost. In this work, we propose a different approach: we learn a differentiable permutation of variables using the Gumbel--Sinkhorn operator and triangularize the instantaneous coefficient matrix of a Structural Vector Autoregressive (SVAR) model in the learned order. This converts acyclicity from a hard constraint into a parameterization and keeps it valid throughout optimization. In doing so, our method enables unified, continuous optimization with gradient-based learning, leading to improved efficiency in time--series causal discovery. Across three real-world benchmarks, our method achieves the best overall performance compared with 12 baselines in both discovery accuracy and efficiency. On the large-scale benchmark, it further demonstrates strong scalability, achieving more than a 6x speedup over competing methods.

Jeffrey T. H. Wong, Cheng Zhang, Louis Mahon et al.

Attention sinks are tokens, often the beginning-of-sequence (BOS) token, that receive disproportionately high attention despite limited semantic relevance. In this work, we identify a class of attention sinks, which we term secondary sinks, that differ fundamentally from the sinks studied in prior works, which we term primary sinks. While prior works have identified that tokens other than BOS can sometimes become sinks, they were found to exhibit properties analogous to the BOS token. Specifically, they emerge at the same layer, persist throughout the network and draw a large amount of attention mass. Whereas, we find the existence of secondary sinks that arise primarily in middle layers and can persist for a variable number of layers, and draw a smaller, but still significant, amount of attention mass. Through extensive experiments across 11 model families, we analyze where these secondary sinks appear, their properties, how they are formed, and their impact on the attention mechanism. Specifically, we show that: (1) these sinks are formed by specific middle-layer MLP modules; these MLPs map token representations to vectors that align with the direction of the primary sink of that layer. (2) The $\ell_2$-norm of these vectors determines the sink score of the secondary sink, and also the number of layers it lasts for, thereby leading to different impacts on the attention mechanisms accordingly. (3) The primary sink weakens in middle layers, coinciding with the emergence of secondary sinks. We observe that in larger-scale models, the location and lifetime of the sinks, together referred to as sink levels, appear in a more deterministic and frequent manner. Specifically, we identify three sink levels in QwQ-32B and six levels in Qwen3-14B. We open-sourced our findings at github.com/JeffreyWong20/Secondary-Attention-Sinks.

Zhiqiang Que, Shuo Liu, Markus Rognlien et al.

This paper introduces a novel optimization framework for deep neural network (DNN) hardware accelerators, enabling the rapid development of customized and automated design flows. More specifically, our approach aims to automate the selection and configuration of low-level optimization techniques, encompassing DNN and FPGA low-level optimizations. We introduce novel optimization and transformation tasks for building design-flow architectures, which are highly customizable and flexible, thereby enhancing the performance and efficiency of DNN accelerators. Our results demonstrate considerable reductions of up to 92\% in DSP usage and 89\% in LUT usage for two networks, while maintaining accuracy and eliminating the need for human effort or domain expertise. In comparison to state-of-the-art approaches, our design achieves higher accuracy and utilizes three times fewer DSP resources, underscoring the advantages of our proposed framework.

Hongxiang Fan, Hao Chen, Liam Castelli et al.

Bayesian Neural Networks (BayesNNs) have demonstrated their capability of providing calibrated prediction for safety-critical applications such as medical imaging and autonomous driving. However, the high algorithmic complexity and the poor hardware performance of BayesNNs hinder their deployment in real-life applications. To bridge this gap, this paper proposes a novel multi-exit Monte-Carlo Dropout (MCD)-based BayesNN that achieves well-calibrated predictions with low algorithmic complexity. To further reduce the barrier to adopting BayesNNs, we propose a transformation framework that can generate FPGA-based accelerators for multi-exit MCD-based BayesNNs. Several novel optimization techniques are introduced to improve hardware performance. Our experiments demonstrate that our auto-generated accelerator achieves higher energy efficiency than CPU, GPU, and other state-of-the-art hardware implementations.

Zehuan Zhang, Mark Chen, He Li et al.

Complex-Valued Neural Networks (CVNNs) have significant advantages in handling tasks that involve complex numbers. However, existing CVNNs are unable to quantify predictive uncertainty. We propose, for the first time, dropout-based Bayesian Complex-Valued Neural Networks (BayesCVNNs) to enable uncertainty quantification for complex-valued applications, exhibiting broad applicability and efficiency for hardware implementation due to modularity. Furthermore, as the dual-part nature of complex values significantly broadens the design space and enables novel configurations based on layer-mixing and part-mixing, we introduce an automated search approach to effectively identify optimal configurations for both real and imaginary components. To facilitate deployment, we present a framework that generates customized FPGA-based accelerators for BayesCVNNs, leveraging a set of optimized building blocks. Experiments demonstrate the best configuration can be effectively found via the automated search, attaining higher performance with lower hardware costs compared with manually crafted models. The optimized accelerators achieve approximately 4.5x and 13x speedups on different models with less than 10% power consumption compared to GPU implementations, and outperform existing work in both algorithm and hardware aspects. Our code is publicly available at: https://github.com/zehuanzhang/BayesCVNN.git.

Zehuan Zhang, Matej Genci, Hongxiang Fan et al.

Accurate and reliable Magnetic Resonance Imaging (MRI) analysis is particularly important for adaptive radiotherapy, a recent medical advance capable of improving cancer diagnosis and treatment. Recent studies have shown that IVIM-NET, a deep neural network (DNN), can achieve high accuracy in MRI analysis, indicating the potential of deep learning to enhance diagnostic capabilities in healthcare. However, IVIM-NET does not provide calibrated uncertainty information needed for reliable and trustworthy predictions in healthcare. Moreover, the expensive computation and memory demands of IVIM-NET reduce hardware performance, hindering widespread adoption in realistic scenarios. To address these challenges, this paper proposes an algorithm-hardware co-optimization flow for high-performance and reliable MRI analysis. At the algorithm level, a transformation design flow is introduced to convert IVIM-NET to a mask-based Bayesian Neural Network (BayesNN), facilitating reliable and efficient uncertainty estimation. At the hardware level, we propose an FPGA-based accelerator with several hardware optimizations, such as mask-zero skipping and operation reordering. Experimental results demonstrate that our co-design approach can satisfy the uncertainty requirements of MRI analysis, while achieving 7.5 times and 32.5 times speedup on an Xilinx VU13P FPGA compared to GPU and CPU implementations with reduced power consumption.

Ziyang Jiao, Ce Guo, Wayne Luk

Causal discovery in time-series data presents a significant computational challenge. Standard algorithms are often prohibitively expensive for datasets with many variables or samples. This study introduces and validates a heuristic approximation of the VarLiNGAM algorithm to address this scalability problem. The standard VarLiNGAM method relies on an iterative search, recalculating statistical dependencies after each step. Our heuristic modifies this procedure by omitting the iterative refinement. This change permits a one-time precomputation of all necessary statistical values. The algorithmic modification reduces the time complexity from $O(m^3n)$ to $O(m^2n + m^3)$ while keeping the space complexity at $O(m^2)$, where $m$ is the number of variables and $n$ is the number of samples. While an approximation, our approach retains VarLiNGAM's essential structure and empirical reliability. On large-scale financial data with up to 400 variables, our algorithm achieves a 7--13x speedup over the standard implementation and a 4.5x speedup over a GPU-accelerated version. Evaluations across medical imaging, web server monitoring, and finance demonstrate the heuristic's robustness and practical scalability. This work offers a validated balance between computational efficiency and discovery quality, making large-scale causal analysis feasible on personal computers.

Chang Sun, Zhiqiang Que, Vladimir Loncar et al.

Neural networks with a latency requirement on the order of microseconds, like the ones used at the CERN Large Hadron Collider, are typically deployed on FPGAs fully unrolled and pipelined. A bottleneck for the deployment of such neural networks is area utilization, which is directly related to the required constant matrix-vector multiplication (CMVM) operations. In this work, we propose an efficient algorithm for implementing CMVM operations with distributed arithmetic (DA) on FPGAs that simultaneously optimizes for area consumption and latency. The algorithm achieves resource reduction similar to state-of-the-art algorithms while being significantly faster to compute. The proposed algorithm is open-sourced and integrated into the \texttt{hls4ml} library, a free and open-source library for running real-time neural network inference on FPGAs. We show that the proposed algorithm can reduce on-chip resources by up to a third for realistic, highly quantized neural networks while simultaneously reducing latency, enabling the implementation of previously infeasible networks.

Zhiqiang Que, Chang Sun, Sudarshan Paramesvaran et al.

Graph Neural Networks (GNNs), particularly Interaction Networks (INs), have shown exceptional performance for jet tagging at the CERN High-Luminosity Large Hadron Collider (HL-LHC). However, their computational complexity and irregular memory access patterns pose significant challenges for deployment on FPGAs in hardware trigger systems, where strict latency and resource constraints apply. In this work, we propose JEDI-linear, a novel GNN architecture with linear computational complexity that eliminates explicit pairwise interactions by leveraging shared transformations and global aggregation. To further enhance hardware efficiency, we introduce fine-grained quantization-aware training with per-parameter bitwidth optimization and employ multiplier-free multiply-accumulate operations via distributed arithmetic. Evaluation results show that our FPGA-based JEDI-linear achieves 3.7 to 11.5 times lower latency, up to 150 times lower initiation interval, and up to 6.2 times lower LUT usage compared to state-of-the-art GNN designs while also delivering higher model accuracy and eliminating the need for DSP blocks entirely. This is the first interaction-based GNN to achieve less than 60~ns latency and currently meets the requirements for use in the HL-LHC CMS Level-1 trigger system. This work advances the next-generation trigger systems by enabling accurate, scalable, and resource-efficient GNN inference in real-time environments. Our open-sourced templates will further support reproducibility and broader adoption across scientific applications.

Zhiwen Mo, Guoyu Li, Hao et al.

Advances in hybrid bonding and packaging have driven growing interest in 3D DRAM-stacked accelerators with higher memory bandwidth and capacity. As LLMs scale to hundreds of billions or trillions of parameters, distributed inference across multiple 3D chips becomes essential. With cross-stack co-design increasingly critical, we propose DeepStack, an accurate and efficient performance model and tool to enable early-stage system-hardware co-design space exploration (DSE) for distributed 3D-stacked AI systems. At the hardware level, DeepStack captures fine-grained 3D memory semantics such as transaction-aware bandwidth, bank activation constraints, buffering limitations, and thermal-power modeling. At the system level, DeepStack incorporates comprehensive parallelization strategies and execution scheduling for distributed LLM inference. With novel modeling techniques such as dual-stage network abstraction and tile-level compute-communication overlap, we achieve up to 100,000x faster runtime over state-of-the-art simulators at comparable accuracy, cross-validated against our in-house 3D designs, NS-3 backend (2.12%), and vLLM serving on 8xB200 GPUs (12.18%). With hierarchical design space search, DeepStack enables efficient exploration over 2.5x10^14 design points spanning 3D-stacked DRAM layers, DRAM vertical connectivity, interconnect, compute-memory allocation, and distributed scheduling. Compared with baseline designs, DeepStack achieves up to 9.5x higher throughput through co-optimized parallelism and 3D architecture search. Our DSE further reveals that batch size drives a more fundamental architectural divide than the prefill/decode distinction, and that parallelism strategy and hardware architecture are tightly coupled -- incomplete schedule search leads to permanently suboptimal silicon irrecoverable by software tuning. We intend to open source DeepStack to support future research.

Hao Mark Chen, Shell Xu Hu, Wayne Luk et al.

Model merging has emerged as a promising approach for multi-task learning (MTL), offering a data-efficient alternative to conventional fine-tuning. However, with the rapid development of the open-source AI ecosystem and the increasing availability of fine-tuned foundation models, existing model merging methods face two key limitations: (i) They are primarily designed for in-house fine-tuned models, making them less adaptable to diverse model sources with partially unknown model and task information, (ii) They struggle to scale effectively when merging numerous model checkpoints. To address these challenges, we formulate model merging as a constrained optimization problem and introduce a novel approach: Frank-Wolfe Merging (FW-Merging). Inspired by Frank-Wolfe optimization, our approach iteratively selects the most relevant model in the pool to minimize a linear approximation of the objective function and then executes a local merging similar to the Frank-Wolfe update. The objective function is designed to capture the desired behavior of the target-merged model, while the fine-tuned candidate models define the constraint set. More importantly, FW-Merging serves as an orthogonal technique for existing merging methods, seamlessly integrating with them to further enhance accuracy performance. Our experiments show that FW-Merging scales across diverse model sources, remaining stable with 16 irrelevant models and improving by 15.3% with 16 relevant models on 20 CV tasks, while maintaining constant memory overhead, unlike the linear overhead of data-informed merging methods. Compared with the state-of-the-art approaches, FW-Merging surpasses the data-free merging method by 32.8% and outperforms the data-informed Adamerging by 8.39% when merging 20 ViT models. Our code is open-sourced at github.com/hmarkc/FW-Merging.

Zhican Wang, Guanghui He, Dantong Liu et al.

3D Gaussian Splatting (3DGS) has recently gained significant attention for high-quality and efficient view synthesis, making it widely adopted in fields such as AR/VR, robotics, and autonomous driving. Despite its impressive algorithmic performance, real-time rendering on resource-constrained devices remains a major challenge due to tight power and area budgets. This paper presents an architecture-algorithm co-design to address these inefficiencies. First, we reveal substantial redundancy caused by repeated computation of common terms/expressions during the conventional rasterization. To resolve this, we propose axis-oriented rasterization, which pre-computes and reuses shared terms along both the X and Y axes through a dedicated hardware design, effectively reducing multiply-and-add (MAC) operations by up to 63%. Second, by identifying the resource and performance inefficiency of the sorting process, we introduce a novel neural sorting approach that predicts order-independent blending weights using an efficient neural network, eliminating the need for costly hardware sorters. A dedicated training framework is also proposed to improve its algorithmic stability. Third, to uniformly support rasterization and neural network inference, we design an efficient reconfigurable processing array that maximizes hardware utilization and throughput. Furthermore, we introduce a $π$-trajectory tile schedule, inspired by Morton encoding and Hilbert curve, to optimize Gaussian reuse and reduce memory access overhead. Comprehensive experiments demonstrate that the proposed design preserves rendering quality while achieving a speedup of $23.4\sim27.8\times$ and energy savings of $28.8\sim51.4\times$ compared to edge GPUs for real-world scenes. We plan to open-source our design to foster further development in this field.

Hao Mark Chen, Liam Castelli, Martin Ferianc et al.

Reliable uncertainty estimation plays a crucial role in various safety-critical applications such as medical diagnosis and autonomous driving. In recent years, Bayesian neural networks (BayesNNs) have gained substantial research and industrial interests due to their capability to make accurate predictions with reliable uncertainty estimation. However, the algorithmic complexity and the resulting hardware performance of BayesNNs hinder their adoption in real-life applications. To bridge this gap, this paper proposes an algorithm and hardware co-design framework that can generate field-programmable gate array (FPGA)-based accelerators for efficient BayesNNs. At the algorithm level, we propose novel multi-exit dropout-based BayesNNs with reduced computational and memory overheads while achieving high accuracy and quality of uncertainty estimation. At the hardware level, this paper introduces a transformation framework that can generate FPGA-based accelerators for the proposed efficient multi-exit BayesNNs. Several optimization techniques such as the mix of spatial and temporal mappings are introduced to reduce resource consumption and improve the overall hardware performance. Comprehensive experiments demonstrate that our approach can achieve higher energy efficiency compared to CPU, GPU, and other state-of-the-art hardware implementations. To support the future development of this research, we have open-sourced our code at: https://github.com/os-hxfan/MCME_FPGA_Acc.git

Patrick Odagiu, Zhiqiang Que, Javier Duarte et al.

Three machine learning models are used to perform jet origin classification. These models are optimized for deployment on a field-programmable gate array device. In this context, we demonstrate how latency and resource consumption scale with the input size and choice of algorithm. Moreover, the models proposed here are designed to work on the type of data and under the foreseen conditions at the CERN LHC during its high-luminosity phase. Through quantization-aware training and efficient synthetization for a specific field programmable gate array, we show that $O(100)$ ns inference of complex architectures such as Deep Sets and Interaction Networks is feasible at a relatively low computational resource cost.

Guoyu Li, Yang Cao, Lucas H L Ng et al.

With network requirements diverging across emerging applications, latency-critical services demand minimal logic delay, while hyperscale training and collectives require sustained line-rate throughput for synchronized bulk transfers. This divergence creates an urgent need for custom network switches tailored to specialized protocols and application-specific traffic patterns. This paper presents SPAC (Switch and Protocol Adaptive Customization), a novel approach that automates the generation of FPGA-based network switches co-optimized for custom protocols and application-specific traffic patterns. SPAC introduces a unified workflow with a domain-specific language (DSL) for protocol-architecture co-design, a library of modular HLS-based adaptive switch components, and a trace-aware Design Space Exploration (DSE) engine. By providing a multi-fidelity simulation stack, SPAC enables rapid identification of Pareto-optimal designs prior to deployment. We demonstrate the efficacy of the domain-specific adaptation of SPAC across a spectrum of real-world scenarios, spanning from latency-sensitive sensor and HFT networks to hyperscale datacenter fabrics. Experimental results show that by tailoring the micro-architecture and protocol to the specific workload, SPAC-generated designs reduce LUT and BRAM usage by 55% and 53%, respectively. Compared to fixed-architecture counterparts, SPAC delivers latency reductions ranging from 7.8% to 38.4% across various tasks while maintaining adequate resource consumption and packet drop rate.

Jeffrey T. H. Wong, Cheng Zhang, Xinye Cao et al.

Large language models have demonstrated remarkable performance; however, their massive parameter counts make deployment highly expensive. Low-rank approximation offers a promising compression solution, yet existing approaches have two main limitations: (1) They focus on minimizing the output error of individual linear layers, without considering the architectural characteristics of Transformers, and (2) they decompose a large weight matrix into two small low-rank matrices. Consequently, these methods often fall short compared to other compression techniques like pruning and quantization, and introduce runtime overhead such as the extra GEMM kernel launches for decomposed small matrices. To address these limitations, we propose $\tt A^\tt 3$, a post-training low-rank approximation framework. $\tt A^\tt 3$ splits a Transformer layer into three functional components, namely $\tt QK$, $\tt OV$, and $\tt MLP$. For each component, $\tt A^\tt 3$ provides an analytical solution that reduces the hidden dimension size inside each component while minimizing the component's functional loss ($\it i.e.$, error in attention scores, attention outputs, and MLP outputs). This approach directly reduces model sizes, KV cache sizes, and FLOPs without introducing any runtime overheads. In addition, it provides a new narrative in advancing the optimization problem from singular linear layer loss optimization toward improved end-to-end performance. Through extensive experiments, we show that $\tt A^\tt 3$ maintains superior performance compared to SoTAs. For example, under the same reduction budget in computation and memory, our low-rank approximated LLaMA 3.1-70B achieves a perplexity of 4.69 on WikiText-2, outperforming the previous SoTA's 7.87 by 3.18. We also demonstrate the versatility of $\tt A^\tt 3$, including KV cache compression, quantization, and mixed-rank assignments for enhanced performance.

Zhiqiang Que, Jose G. F. Coutinho, Ce Guo et al.

This paper presents a unified framework for codifying and automating optimization strategies to efficiently deploy deep neural networks (DNNs) on resource-constrained hardware, such as FPGAs, while maintaining high performance, accuracy, and resource efficiency. Deploying DNNs on such platforms involves addressing the significant challenge of balancing performance, resource usage (e.g., DSPs and LUTs), and inference accuracy, which often requires extensive manual effort and domain expertise. Our novel approach addresses two core key issues: (i)~encoding custom optimization strategies and (ii)~enabling cross-stage optimization search. In particular, our proposed framework seamlessly integrates programmatic DNN optimization techniques with high-level synthesis (HLS)-based metaprogramming, leveraging advanced design space exploration (DSE) strategies like Bayesian optimization to automate both top-down and bottom-up design flows. Hence, we reduce the need for manual intervention and domain expertise. In addition, the framework introduces customizable optimization, transformation, and control blocks to enhance DNN accelerator performance and resource efficiency. Experimental results demonstrate up to a 92\% DSP and 89\% LUT usage reduction for select networks, while preserving accuracy, along with a 15.6-fold reduction in optimization time compared to grid search. These results highlight the potential for automating the generation of resource-efficient DNN accelerator designs with minimum effort.

Gene Yu, Ce Guo, Wayne Luk

Agent-Based Model (ABM) validation is crucial as it helps ensuring the reliability of simulations, and causal discovery has become a powerful tool in this context. However, current causal discovery methods often face accuracy and robustness challenges when applied to complex and noisy time series data, which is typical in ABM scenarios. This study addresses these issues by proposing a Robust Cross-Validation (RCV) approach to enhance causal structure learning for ABM validation. We develop RCV-VarLiNGAM and RCV-PCMCI, novel extensions of two prominent causal discovery algorithms. These aim to reduce the impact of noise better and give more reliable causal relation results, even with high-dimensional, time-dependent data. The proposed approach is then integrated into an enhanced ABM validation framework, which is designed to handle diverse data and model structures. The approach is evaluated using synthetic datasets and a complex simulated fMRI dataset. The results demonstrate greater reliability in causal structure identification. The study examines how various characteristics of datasets affect the performance of established causal discovery methods. These characteristics include linearity, noise distribution, stationarity, and causal structure density. This analysis is then extended to the RCV method to see how it compares in these different situations. This examination helps confirm whether the results are consistent with existing literature and also reveals the strengths and weaknesses of the novel approaches. By tackling key methodological challenges, the study aims to enhance ABM validation with a more resilient valuation framework presented. These improvements increase the reliability of model-driven decision making processes in complex systems analysis.

Guanxi Lu, Hao Mark Chen, Zhiqiang Que et al.

Large language models (LLMs) have shown promising performance across various tasks. However, their autoregressive decoding process poses significant challenges for efficient deployment on existing AI hardware. Quantization alleviates memory and compute pressure by compressing weights, activations, and KV caches to low precisions while preserving generation quality. However, existing quantization frameworks typically focus on perplexity or classification accuracy, often omitting critical trustworthiness metrics. This gap introduces risks when applying quantized LLMs to downstream high-stakes domains such as finance and healthcare. In this work, we systematically investigate the impact of quantization on four trustworthiness metrics (adversarial robustness, fairness, machine ethics, and out-of-distribution robustness) and identify the instability across compression ratios and quantization methods. Building on these observations, we develop a novel precision-ensemble voting approach that leverages predictions from mixed-precision variants of the same model and consistently improves performance by up to $5.8\%$ on trustworthiness metrics. Our results highlight the importance of considering trustworthiness when developing model compression techniques and point to research opportunities at the intersection of compression and trustworthiness for safety-critical applications.

Lauri Laatu, Chang Sun, Arianna Cox et al.

We present the first sub-microsecond transformer implementation on an FPGA achieving competitive performance for state-of-the-art high-energy physics benchmarks. Transformers have shown exceptional performance on multiple tasks in modern machine learning applications, including jet tagging at the CERN Large Hadron Collider (LHC). However, their computational complexity prohibits use in real-time applications, such as the hardware trigger system of the collider experiments up until now. In this work, we demonstrate the first application of transformers for jet tagging on FPGAs, achieving $\mathcal{O}(100)$ nanosecond latency with superior performance compared to alternative baseline models. We leverage high-granularity quantization and distributed arithmetic optimization to fit the entire transformer model on a single FPGA, achieving the required throughput and latency. Furthermore, we add multi-head attention and linear attention support to hls4ml, making our work accessible to the broader fast machine learning community. This work advances the next-generation trigger systems for the High Luminosity LHC, enabling the use of transformers for real-time applications in high-energy physics and beyond.

Hao Mark Chen, Zhiwen Mo, Guanxi Lu et al.

Deploying agentic AI on edge devices is crucial for privacy and responsiveness, but memory constraints typically relegate these systems to smaller Large Language Models (LLMs) with inferior reasoning capabilities. Test-Time Scaling (TTS) can bridge this reasoning gap by dedicating more compute during inference, but existing methods incur prohibitive overhead on edge hardware. To overcome this, we introduce FlashTTS, a serving system that makes TTS practical for memory-constrained LLM reasoning. FlashTTS introduces three synergistic optimizations: (i) Speculative Beam Extension to mitigate system stragglers from irregular reasoning paths; (ii) Asymmetric Multi-Model Memory Allocation to dynamically balance memory between generation and verification; and (iii) Dynamic Prefix-Aware Scheduling to maximize KV-cache reuse. Built as a plug-and-play library for vLLM, FlashTTS enables edge LLMs on a single consumer GPU (24 GB) to match the accuracy and latency of large cloud models. Our evaluation demonstrates that FlashTTS achieves an average 2.2x higher goodput and reduces latency by 38%-68% compared to a vLLM baseline, paving the way for democratized, high-performance agentic AI on edge devices.

Zehuan Zhang, Hongxiang Fan, Hao Mark Chen et al.

The increasing deployment of artificial intelligence (AI) for critical decision-making amplifies the necessity for trustworthy AI, where uncertainty estimation plays a pivotal role in ensuring trustworthiness. Dropout-based Bayesian Neural Networks (BayesNNs) are prominent in this field, offering reliable uncertainty estimates. Despite their effectiveness, existing dropout-based BayesNNs typically employ a uniform dropout design across different layers, leading to suboptimal performance. Moreover, as diverse applications require tailored dropout strategies for optimal performance, manually optimizing dropout configurations for various applications is both error-prone and labor-intensive. To address these challenges, this paper proposes a novel neural dropout search framework that automatically optimizes both the dropout-based BayesNNs and their hardware implementations on FPGA. We leverage one-shot supernet training with an evolutionary algorithm for efficient dropout optimization. A layer-wise dropout search space is introduced to enable the automatic design of dropout-based BayesNNs with heterogeneous dropout configurations. Extensive experiments demonstrate that our proposed framework can effectively find design configurations on the Pareto frontier. Compared to manually-designed dropout-based BayesNNs on GPU, our search approach produces FPGA designs that can achieve up to 33X higher energy efficiency. Compared to state-of-the-art FPGA designs of BayesNN, the solutions from our approach can achieve higher algorithmic performance and energy efficiency.

Hongxiang Fan, Martin Ferianc, Zhiqiang Que et al.

Recent advances in algorithm-hardware co-design for deep neural networks (DNNs) have demonstrated their potential in automatically designing neural architectures and hardware designs. Nevertheless, it is still a challenging optimization problem due to the expensive training cost and the time-consuming hardware implementation, which makes the exploration on the vast design space of neural architecture and hardware design intractable. In this paper, we demonstrate that our proposed approach is capable of locating designs on the Pareto frontier. This capability is enabled by a novel three-phase co-design framework, with the following new features: (a) decoupling DNN training from the design space exploration of hardware architecture and neural architecture, (b) providing a hardware-friendly neural architecture space by considering hardware characteristics in constructing the search cells, (c) adopting Gaussian process to predict accuracy, latency and power consumption to avoid time-consuming synthesis and place-and-route processes. In comparison with the manually-designed ResNet101, InceptionV2 and MobileNetV2, we can achieve up to 5% higher accuracy with up to 3x speed up on the ImageNet dataset. Compared with other state-of-the-art co-design frameworks, our found network and hardware configuration can achieve 2% ~ 6% higher accuracy, 2x ~ 26x smaller latency and 8.5x higher energy efficiency.

Zhiqiang Que, Erwei Wang, Umar Marikar et al.

This paper presents novel reconfigurable architectures for reducing the latency of recurrent neural networks (RNNs) that are used for detecting gravitational waves. Gravitational interferometers such as the LIGO detectors capture cosmic events such as black hole mergers which happen at unknown times and of varying durations, producing time-series data. We have developed a new architecture capable of accelerating RNN inference for analyzing time-series data from LIGO detectors. This architecture is based on optimizing the initiation intervals (II) in a multi-layer LSTM (Long Short-Term Memory) network, by identifying appropriate reuse factors for each layer. A customizable template for this architecture has been designed, which enables the generation of low-latency FPGA designs with efficient resource utilization using high-level synthesis tools. The proposed approach has been evaluated based on two LSTM models, targeting a ZYNQ 7045 FPGA and a U250 FPGA. Experimental results show that with balanced II, the number of DSPs can be reduced up to 42% while achieving the same IIs. When compared to other FPGA-based LSTM designs, our design can achieve about 4.92 to 12.4 times lower latency.

Martin Ferianc, Zhiqiang Que, Hongxiang Fan et al.

Neural networks have demonstrated their outstanding performance in a wide range of tasks. Specifically recurrent architectures based on long-short term memory (LSTM) cells have manifested excellent capability to model time dependencies in real-world data. However, standard recurrent architectures cannot estimate their uncertainty which is essential for safety-critical applications such as in medicine. In contrast, Bayesian recurrent neural networks (RNNs) are able to provide uncertainty estimation with improved accuracy. Nonetheless, Bayesian RNNs are computationally and memory demanding, which limits their practicality despite their advantages. To address this issue, we propose an FPGA-based hardware design to accelerate Bayesian LSTM-based RNNs. To further improve the overall algorithmic-hardware performance, a co-design framework is proposed to explore the most fitting algorithmic-hardware configurations for Bayesian RNNs. We conduct extensive experiments on healthcare applications to demonstrate the improvement of our design and the effectiveness of our framework. Compared with GPU implementation, our FPGA-based design can achieve up to 10 times speedup with nearly 106 times higher energy efficiency. To the best of our knowledge, this is the first work targeting acceleration of Bayesian RNNs on FPGAs.

Hongxiang Fan, Martin Ferianc, Miguel Rodrigues et al.

Neural networks (NNs) have demonstrated their potential in a wide range of applications such as image recognition, decision making or recommendation systems. However, standard NNs are unable to capture their model uncertainty which is crucial for many safety-critical applications including healthcare and autonomous vehicles. In comparison, Bayesian neural networks (BNNs) are able to express uncertainty in their prediction via a mathematical grounding. Nevertheless, BNNs have not been as widely used in industrial practice, mainly because of their expensive computational cost and limited hardware performance. This work proposes a novel FPGA-based hardware architecture to accelerate BNNs inferred through Monte Carlo Dropout. Compared with other state-of-the-art BNN accelerators, the proposed accelerator can achieve up to 4 times higher energy efficiency and 9 times better compute efficiency. Considering partial Bayesian inference, an automatic framework is proposed, which explores the trade-off between hardware and algorithmic performance. Extensive experiments are conducted to demonstrate that our proposed framework can effectively find the optimal points in the design space.

Seyedeh Niusha Alavi Foumani, Ce Guo, Wayne Luk

In this work, we present a hardware compatible neural network training algorithm in which we used alternating direction method of multipliers (ADMM) and iterative least-square methods. The motive behind this approach was to conduct a method of training neural networks that is scalable and can be parallelised. These characteristics make this algorithm suitable for hardware implementation. We have achieved 6.9\% and 6.8\% better accuracy comparing to SGD and Adam respectively, with a four-layer neural network with hidden size of 28 on HIGGS dataset. Likewise, we could observe 21.0\% and 2.2\% accuracy improvement comparing to SGD and Adam respectively, on IRIS dataset with a three-layer neural network with hidden size of 8. This is while the use of matrix inversion, which is challenging for hardware implementation, is avoided in this method. We assessed the impact of avoiding matrix inversion on ADMM accuracy and we observed that we can safely replace matrix inversion with iterative least-square methods and maintain the desired performance. Also, the computational complexity of the implemented method is polynomial regarding dimensions of the input dataset and hidden size of the network.

Seyedeh Niusha Alavi Foumani, Ce Guo, Wayne Luk

In this project, we have successfully designed, implemented, deployed and tested a novel FPGA accelerated algorithm for neural network training. The algorithm itself was developed in an independent study option. This training method is based on Alternating Direction Method of Multipliers algorithm, which has strong parallel characteristics and avoids procedures such as matrix inversion that are problematic in hardware designs by employing LSMR. As an intermediate stage, we fully implemented the ADMM-LSMR method in C language for feed-forward neural networks with a flexible number of layers and hidden size. We demonstrated that the method can operate with fixed-point arithmetic without compromising the accuracy. Next, we devised an FPGA accelerated version of the algorithm using Intel FPGA SDK for OpenCL and performed extensive optimisation stages followed by successful deployment of the program on an Intel Arria 10 GX FPGA. The FPGA accelerated program showed up to 6 times speed up comparing to equivalent CPU implementation while achieving promising accuracy.

Yang Chu, Wayne Luk, Dan Goodman

The acoustic cues used by humans and other animals to localise sounds are subtle, and change during and after development. This means that we need to constantly relearn or recalibrate the auditory spatial map throughout our lifetimes. This is often thought of as a "supervised" learning process where a "teacher" (for example, a parent, or your visual system) tells you whether or not you guessed the location correctly, and you use this information to update your map. However, there is not always an obvious teacher (for example in babies or blind people). Using computational models, we showed that approximate feedback from a simple innate circuit, such as that can distinguish left from right (e.g. the auditory orienting response), is sufficient to learn an accurate full-range spatial auditory map. Moreover, using this mechanism in addition to supervised learning can more robustly maintain the adaptive neural representation. We find several possible neural mechanisms that could underlie this type of learning, and hypothesise that multiple mechanisms may be present and interact with each other. We conclude that when studying spatial hearing, we should not assume that the only source of learning is from the visual system or other supervisory signal. Further study of the proposed mechanisms could allow us to design better rehabilitation programmes to accelerate relearning/recalibration of spatial maps.

Ringo S. W. Chu, Ho-Cheung Ng, Xiwei Wang et al.

Hyperspectral images (HSIs) can distinguish materials with high number of spectral bands, which is widely adopted in remote sensing applications and benefits in high accuracy land cover classifications. However, HSIs processing are tangled with the problem of high dimensionality and limited amount of labelled data. To address these challenges, this paper proposes a deep learning architecture using three dimensional convolutional neural networks with spectral partitioning to perform effective feature extraction. We conduct experiments using Indian Pines and Salinas scenes acquired by NASA Airborne Visible/Infra-Red Imaging Spectrometer. In comparison to prior results, our architecture shows competitive performance for classification results over current methods.

Shuanglong Liu, Ringo S. W. Chu, Xiwei Wang et al.

Hyperspectral image (HSI) classification has been widely adopted in applications involving remote sensing imagery analysis which require high classification accuracy and real-time processing speed. Methods based on Convolutional neural networks (CNNs) have been proven to achieve state-of-the-art accuracy in classifying HSIs. However, CNN models are often too computationally intensive to achieve real-time response due to the high dimensional nature of HSI, compared to traditional methods such as Support Vector Machines (SVMs). Besides, previous CNN models used in HSI are not specially designed for efficient implementation on embedded devices such as FPGAs. This paper proposes a novel CNN-based algorithm for HSI classification which takes into account hardware efficiency. A customized architecture which enables the proposed algorithm to be mapped effectively onto FPGA resources is then proposed to support real-time on-board classification with low power consumption. Implementation results show that our proposed accelerator on a Xilinx Zynq 706 FPGA board achieves more than 70x faster than an Intel 8-core Xeon CPU and 3x faster than an NVIDIA GeForce 1080 GPU. Compared to previous SVM-based FPGA accelerators, we achieve comparable processing speed but provide a much higher classification accuracy.

Erwei Wang, James J. Davis, Ruizhe Zhao et al.

Deep neural networks have proven to be particularly effective in visual and audio recognition tasks. Existing models tend to be computationally expensive and memory intensive, however, and so methods for hardware-oriented approximation have become a hot topic. Research has shown that custom hardware-based neural network accelerators can surpass their general-purpose processor equivalents in terms of both throughput and energy efficiency. Application-tailored accelerators, when co-designed with approximation-based network training methods, transform large, dense and computationally expensive networks into small, sparse and hardware-efficient alternatives, increasing the feasibility of network deployment. In this article, we provide a comprehensive evaluation of approximation methods for high-performance network inference along with in-depth discussion of their effectiveness for custom hardware implementation. We also include proposals for future research based on a thorough analysis of current trends. This article represents the first survey providing detailed comparisons of custom hardware accelerators featuring approximation for both convolutional and recurrent neural networks, through which we hope to inspire exciting new developments in the field.

Ruizhe Zhao, Wayne Luk

Efficient inference of Convolutional Neural Networks is a thriving topic recently. It is desirable to achieve the maximal test accuracy under given inference budget constraints when deploying a pre-trained model. Network pruning is a commonly used technique while it may produce irregular sparse models that can hardly gain actual speed-up. Group convolution is a promising pruning target due to its regular structure; however, incorporating such structure into the pruning procedure is challenging. It is because structural constraints are hard to describe and can make pruning intractable to solve. The need for configuring group convolution architecture, i.e., the number of groups, that maximises test accuracy also increases difficulty. This paper presents an efficient method to address this challenge. We formulate group convolution pruning as finding the optimal channel permutation to impose structural constraints and solve it efficiently by heuristics. We also apply local search to exploring group configuration based on estimated pruning cost to maximise test accuracy. Compared to prior work, results show that our method produces competitive group convolution models for various tasks within a shorter pruning period and enables rapid group configuration exploration subject to inference budget constraints.

Ruizhe Zhao, Ho-Cheung Ng, Wayne Luk et al.

FPGA becomes a popular technology for implementing Convolutional Neural Network (CNN) in recent years. Most CNN applications on FPGA are domain-specific, e.g., detecting objects from specific categories, in which commonly-used CNN models pre-trained on general datasets may not be efficient enough. This paper presents TuRF, an end-to-end CNN acceleration framework to efficiently deploy domain-specific applications on FPGA by transfer learning that adapts pre-trained models to specific domains, replacing standard convolution layers with efficient convolution blocks, and applying layer fusion to enhance hardware design performance. We evaluate TuRF by deploying a pre-trained VGG-16 model for a domain-specific image recognition task onto a Stratix V FPGA. Results show that designs generated by TuRF achieve better performance than prior methods for the original VGG-16 and ResNet-50 models, while for the optimised VGG-16 model TuRF designs are more accurate and easier to process.