The rapid development of large language models (LLM) has greatly enhanced everyday applications. While many FPGA-based accelerators, with flexibility for fine-grained data control, exhibit superior speed and energy efficiency compared to GPUs, recent GPU-specific optimizations have diminished this advantage. When limited to arithmetic-based computation, FPGAs often underperform GPUs due to their comparatively fewer computational resources. To address this challenge, we exploit a key advantage of FPGAs over GPUs: abundant distributed on-chip memory embedded among computational units. We believe that shifting LLM inference from arithmetic-based to memory-based computations through table lookups can improve the efficiency on FPGAs to compete with GPUs. However, existing methods are inefficient or unable to scale and deploy language models due to algorithm and architecture design limitations. This paper introduces LUT-LLM1, the first FPGA accelerator that deploys 1B+ language model with memory-based computation, leveraging vector quantization. We construct a performance model, evaluate multiple quantization schemes, and identify the activation-weight vector co-quantization as the most effective approach. To support this scheme, LUT-LLM features (1) a bandwidth-aware parallel centroid search to reduce decoding latency, (2) efficient 2D table lookups, and (3) a spatial-temporal hybrid design to reduce data caching for a higher throughput table lookup. We develop a training recipe that converts existing models to support table lookups with high accuracy and prototype LUT-LLM for Qwen 3 1.7B model on the AMD V80 FPGA, reducing arithmetic operations by 4× and achieving a 1.10∼3.29× faster generation speed and a 3.05∼6.60× higher energy efficiency than GPUs.

Large language models (LLMs) based on transformer architecture have shown outstanding performance across numerous real-world tasks. However, the autoregressive nature of these models makes the inference process slow and costly. Speculative decoding has emerged as a promising solution, leveraging a smaller auxiliary model to draft future tokens, which are then validated simultaneously by the larger model, achieving a speed-up of 1-2x. Although speculative decoding matches the same distribution as multinomial sampling, multinomial sampling itself is prone to suboptimal outputs, where as beam sampling is widely recognized for producing higher-quality results by maintaining multiple candidate sequences at each step. This paper explores the novel integration of speculative decoding with beam sampling. However, there are four key challenges: (1) how to generate multiple sequences from the larger model's distribution given drafts sequences from the small model; (2) how to dynamically optimize the number of beams to balance efficiency and accuracy; (3) how to efficiently verify the multiple drafts in parallel; and (4) how to address the extra memory costs inherent in beam sampling. To address these challenges, we propose dynamic-width speculative beam decoding (DSBD). Specifically, we first introduce a novel draft and verification scheme that generates multiple sequences following the large model's distribution based on beam sampling trajectories from the small model. Then, we introduce an adaptive mechanism to dynamically tune the number of beams based on the context, optimizing efficiency and effectiveness. Besides, we extend tree-based parallel verification to handle multiple trees simultaneously, accelerating the verification process. Finally, we illustrate a simple modification to our algorithm to mitigate the memory overhead of beam sampling. Experimental results show that our approach achieves a 1.5-1.9x speed-up and1.8-2.5x lower energy consumption compared to beam sampling, with no loss in downstream performance. Moreover, it can produce significantly higher-quality outputs than speculative decoding, while maintaining similar time, memory, and energy costs. In summary, our method offers a more efficient and effective inference process for LLMs.

This work presents a novel monolithic 3D (M3D) FPGA architecture that leverages stackable back-end-of-line (BEOL) transistors to implement configuration memory and pass gates, significantly improving area, latency, and power efficiency. By integrating n-type (W -doped In2O3) and p-type (SnO) amorphous oxide semiconductor (AOS) transistors in the BEOL, Si SRAM configuration bits are substituted with a less leaky equivalent that can be programmed at logic-compatible voltages. BEOL-compatible AOS transistors are currently under extensive research and development in the device community, with investment by leading foundries, from which reported data is used to develop robust physics-based models in TCAD that enable circuit design. The use of AOS pass gates reduces the overhead of reconfigurable circuits by mapping FPGA switch block (SB) and connection block (CB) matrices above configurable logic blocks (CLBs), thereby increasing the proximity of logic elements and reducing latency. By interfacing with the latest Verilog-to-Routing (VTR) suite, an AOS-based M3D FPGA implemented in 7 nm technology is demonstrated with $3.4 \times$ lower area-time squared product ($\mathbf{A T}^{2}$), 27% lower critical path latency, and 26% lower reconfigurable routing block power on benchmarks including hyperdimensional computing and large language models (LLMs).

The rise of deep neural networks (DNNs) has driven an increased demand for computing power and memory. Modern DNNs exhibit high data volume variation (HDV) across tasks, which poses challenges for FPGA acceleration: conventional accelerators rely on fixed execution patterns (dataflow or sequential) that can lead to pipeline stalls or necessitate frequent off-chip memory accesses. To address these challenges, we introduce the Inter-Task Auto- Reconfigurable Accelerator (InTAR)11https://github.com/OswaldHe/lnTAR, a novel accelerator design methodology for HDV applications on FPGAs. InTAR combines the high computational efficiency of sequential execution with the reduced off-chip memory overhead of dataflow execution. It switches execution patterns automatically with a static schedule determined before circuit design based on resource constraints and problem sizes. Unlike previous reconfigurable accelerators, InTAR encodes reconfiguration schedules during circuit design, allowing model-specific optimizations that allocate only the necessary logic and interconnects. Thus, InTAR achieves a high clock frequency with fewer resources and low reconfiguration time. Furthermore, InTAR supports high-level tools such as HLS for fast design generation. We implement a set of multi-task HDV DNN kernels using InTAR. Compared with dataflow and sequential accelerators, InTAR exhibits 1.8× and 7.1× speedups correspondingly. Moreover, we extend InTAR to GPT-2 medium as a more complex example, which is 3.65 ~39.14× faster and a 1.72 ~ 10.44× more DSP efficient than SoTA accelerators (Allo and DFX) on FPGAs. Additionally, this design demonstrates 1.66 ~ 7.17× better power efficiency than GPUs.

In FPGAs, high communication latency in multi-die chips has driven the integration of hardened networks-on-chip (NoCs) in commercial devices. However, for programming FPGAs with high-level synthesis (HLS), existing tools only provide low-level cumbersome abstractions, and only work for offloading memory accesses. Furthermore, these abstractions remain inaccessible to programmers due to their reliance on placement knowledge. While automatically leveraging the NoC without manual intervention is ideal, it poses several challenges: 1. Managing the trade-off in resource utilization between the hard NoC and the Programmable Logic (PL). 2. Allocating limited hard NoC resources between different communication in the designs. 3. Aligning hard NoC and PL placement even though the actual PL placement cannot be determined beforehand. We address these challenges by developing NoH, the first HLS flow that automates hard NoC offloading. First, we develop a formal NoC-aware placement algorithm that leverages integer linear programming (ILP) and considers the first two challenges for offloading external memory accesses and latency-insensitive communication between modules. Then, we arrange the ports synergistically with PL modules via a port-affinity model that approximates the PL placement. Finally, NoH is integrated into an end-to-end HLS flow and evaluated on 4 workloads with diverse communication patterns. NoH gains 20% FPGA frequency over AMD tools by leveraging the hard NoC. Compared to AutoBridge [1], a recent high-level physical synthesis technique that optimizes frequency but does not consider the hard NoC, NoH never fails place-and-route by offloading inter-die crossings (AutoBridge fails in 31% of workload configurations tested) and is faster (6%) for the rest.

The rise of deep neural networks (DNNs) has driven a boom in AI services, accompanied by increasing demands for computing power and memory. One of the most widely used DNNs is the transformer model. Unlike convolutional neural networks (CNNs), transformers involve more complex computations and generate larger intermediate data, which pose significant challenges for computation efficiency and off-chip memory access overhead in large language model (LLM) accelerator designs. Unfortunately, existing LLM accelerators have fixed execution patterns (either dataflow or sequential), resulting in either low computational efficiency due to pipeline stalls or frequent off-chip memory accesses to manage large intermediate results. This ultimately leads to high single-batch inference latency.

Implementing HLS accelerators on large-scale multi-die FPGAs presents significant challenges. To address this, researchers have proposed High-Level Physical Synthesis (HLPS), which co-optimizes high-level synthesis and physical design to improve achievable frequency. However, existing HLPS techniques suffer from unstable and inconsistent quality of results (QoRs), largely due to the vast number of parameters that need to be selected by the user in an ad-hoc way. As a result, achieving satisfactory solutions still requires substantial manual effort and expertise in low-level circuit design.We propose a robust and practical design space exploration (DSE) framework that enhances the reliability and QoRs of HLPS by automating the iterative parameter tuning process. Informed by metrics extracted from physical implementation outcomes, the framework applies tailored heuristics to refine HLPS parameters, enabling consistent and automated timing closure. In evaluations with large-scale, real-world designs implemented on representative multi-die devices, our framework achieves an average frequency of 311.06 MHz, reaching 2.42× the frequency of the AMD Vitis/Vivado toolchain (128.48 MHz) and 1.67× that of the leading academic solutions (186.21 MHz).

Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42× speed-up and consumes 1.54× less energy than conventional speculative decoding methods. These results highlight MTAD’s ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.

Zifan He, Yingqi Cao, Zongyue Qin, Neha Prakriya, Yizhou Sun, Jason Cong. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Over the past decade, much progress has been made to advance the acceleration of sparse linear operators such as SpMM and SpMV on FPGAs. Nevertheless, few works have attempted to address sparse triangular solver (SpTRSV) acceleration, and the performance boost is limited. SpTRSV is an elementary linear operator for many numerical methods, such as the least-square method. These methods, among others, are widely used in various areas, such as physical simulation and signal processing. Therefore, accelerating SpTRSV is crucial. However, many challenges impede accelerating SpTRSV, including (1) resolving dependencies between elements during forward or backward substitutions, (2) random access and unbalanced workloads across memory channels due to sparsity, (3) latency incurred by off-chip memory access for large matrices or vectors, and (4) data reuse for an unpredictable data sharing pattern. To address these issues, we have designed LevelST, the first FPGA accelerator leveraging high bandwidth memory (HBM) for solving sparse triangular systems. LevelST features (1) algorithm-hardware co-design of stream-based dependency resolution with reduced off-chip data movement, (2) resource sharing that improves resource utilization to scale up the architecture, (3) index modulo scheduling to balance workload, and (4) selective data prefetching from off-chip memory. LevelST is prototyped on an AMD Xilinx U280 HBM FPGA and evaluated with 16 sparse triangular matrices. Compared with the NVIDIA V100 and RTX 3060 GPUs over the cuSPARSE library, LevelST achieves a 2.65x speedup and 9.82x higher energy efficiency than the best of the V100 GPU and RTX 3060 GPU. The code is released on https://github.com/OswaldHe/LevelST (DOI: https://doi.org/10.5281/zenodo.10463345).

Accompanied by improving accessibility of data and storage and the invention of the blockchain economy, advantages of the peer-to-peer network in privacy and efficiency over server-client systems have become more significant recently. One of the most popular applications of peer-to-peer networks is file transferring, where each peer stores a partial segment of a file and the requester can aggregate each piece into a single file when downloading it. To optimize the performance of file transmission, one goal is to increase the searching speed for the file location over the network. In this paper, we analyze the server-mediated peer-to-peer networks, which is an uncommon architecture mentioned in previous works but can potentially improve the search speed, and minimize the average number of hops during flooding search.