Modern large language models rely on attention mechanisms that attend to all tokens in a sequence, resulting in quadratic computational complexity that limits scalability. While sparse attention reduces compute and memory requirements by attending to only important tokens, implementing these techniques presents significant challenges due to the complexity of combining static and dynamic sparse patterns and optimizing key-value (KV) cache management.

Recent advancements in large language models (LLMs) boasting billions of parameters have generated a significant demand for efficient deployment in inference workloads. While hardware accelerators for Transformer-based models have been extensively studied, the majority of existing approaches rely on temporal architectures that reuse hardware units for different network layers and operators. However, these methods often encounter challenges in achieving low latency due to considerable memory access overhead. This article investigates the feasibility and potential of model-specific spatial acceleration for LLM inference on field-programmable gate arrays (FPGAs). Our approach involves the specialization of distinct hardware units for specific operators or layers, facilitating direct communication between them through a dataflow architecture while minimizing off-chip memory accesses. We introduce a comprehensive analytical model for estimating the performance of a spatial LLM accelerator, taking into account the on-chip compute and memory resources available on an FPGA. This model can be extended to multi-FPGA settings for distributed inference. Through our analysis, we can identify the most effective parallelization and buffering schemes for the accelerator and, crucially, determine the scenarios in which FPGA-based spatial acceleration can outperform its GPU-based counterpart. To enable more productive implementations of an LLM model on FPGAs, we further provide a library of high-level synthesis (HLS) kernels that are composable and reusable. This library will be made available as open-source. To validate the effectiveness of both our analytical model and HLS library, we have implemented Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformers (GPT2) on an AMD Xilinx Alveo U280 FPGA device. Experimental results demonstrate our approach can achieve up to 13.4× speedup when compared to previous FPGA-based accelerators for the BERT model. For GPT generative inference, we attain a 2.2× speedup compared to Design for Excellence, an FPGA overlay, in the prefill stage, while achieving a 1.9× speedup and a 5.7× improvement in energy efficiency compared to the NVIDIA A100 GPU in the decode stage.

Recent advancements in large language models (LLMs) have generated significant demands for efficient deployment in inference workloads. Most existing approaches rely on temporal architectures that reuse hardware units for different network layers and operators. However, these methods often encounter challenges in achieving low latency due to considerable memory access overhead. This paper investigates the feasibility and potential of model-specific spatial acceleration for LLM inference on FPGAs. Our approach involves the specialization of distinct hardware units for specific operators, facilitating direct communication between them through a dataflow architecture while minimizing off-chip memory accesses. We introduce a comprehensive analytical model for estimating the performance of a spatial LLM accelerator, considering the available hardware resources on an FPGA. Through our analysis, we can determine the scenarios in which FPGA-based spatial acceleration can outperform its GPU-based counterpart. To enable more productive implementations of an LLM model on FPGAs, we further provide a library of high-level synthesis (HLS) kernels that are composable and reusable. To validate the effectiveness of both our analytical model and HLS library, we have implemented BERT and GPT2 on an AMD Alveo U280 FPGA. Experimental results demonstrate our approach can achieve up to 16.1x speedup when compared to previous FPGA-based accelerators for the BERT model. For GPT generative inference, we attain a 2.2x speedup compared to DFX, an FPGA overlay, in the prefill stage, while achieving a 1.9x speedup and a 5.7x improvement in energy efficiency compared to the NVIDIA A100 GPU in the decode stage.

The increasing complexity of large-scale FPGA accelerators poses significant challenges in achieving high performance while maintaining design productivity. High-level synthesis (HLS) has been adopted as a solution, but the mismatch between the high-level description and the physical layout often leads to suboptimal operating frequency. Although existing proposals for high-level physical synthesis, which use coarse-grained design partitioning, floorplanning, and pipelining to improve frequency, have gained traction, they lack a framework enabling (1) pipelining of real-world designs at arbitrary hierarchical levels, (2) integration of HLS blocks, vendor IPs, and handcrafted RTL designs, (3) portability to emerging new target FPGA devices, and (4) extensibility for the easy implementation of new design optimization tools.

Special-purpose hardware accelerators are increasingly pivotal for sustaining performance improvements in emerging applications, especially as the benefits of technology scaling continue to diminish. However, designers currently lack effective tools and methodologies to construct complex, high-performance accelerator architectures in a productive manner. Existing high-level synthesis (HLS) tools often require intrusive sourcelevel changes to attain satisfactory quality of results. Despite the introduction of several new accelerator design languages (ADLs) aiming to enhance or replace HLS, their advantages are more evident in relatively simple applications with a single kernel. Existing ADLs prove less effective for realistic hierarchical designs with multiple kernels, even if the design hierarchy is flattened. In this paper, we introduce Allo, a composable programming model for efficient spatial accelerator design. Allo decouples hardware customizations, including compute, memory, communication, and data type from algorithm specification, and encapsulates them as a set of customization primitives. Allo preserves the hierarchical structure of an input program by combining customizations from different functions in a bottomup, type-safe manner. This approach facilitates holistic optimizations that span across function boundaries. We conduct comprehensive experiments on commonly-used HLS benchmarks and several realistic deep learning models. Our evaluation shows that Allo can outperform state-of-the-art HLS tools and ADLs on all test cases in the PolyBench. For the GPT2 model, the inference latency of the Allo generated accelerator is 1.7× faster than the NVIDIA A100 GPU with 5.4× higher energy efficiency, demonstrating the capability of Allo to handle large-scale designs.

The past decade has witnessed increasing adoption of high-level synthesis (HLS) to implement specialized hardware accelerators targeting either FPGAs or ASICs. However, current HLS programming models entangle algorithm specifications with hardware customization techniques, which lowers both the productivity and portability of the accelerator design. To tackle this problem, recent efforts such as HeteroCL propose to decouple algorithm definition from essential hardware customization techniques in compute, data type, and memory, increasing productivity, portability, and performance.

To achieve high performance with FPGA-equipped heterogeneous compute systems, it is crucial to co-optimize data placement and compute scheduling to maximize data reuse and bandwidth utilization for both on- and off-chip memory accesses. However, optimizing the data placement for FPGA accelerators is a complex task. One must acquire in-depth knowledge of the target FPGA device and its associated memory system in order to apply a set of advanced optimizations. Even with the latest high-level synthesis (HLS) tools, programmers often have to insert many low-level vendor-specific pragmas and substantially restructure the algorithmic code so that the right data are accessed at the right loop level using the right communication schemes. These code changes can significantly compromise the composability and portability of the original program. To address these challenges, we propose HeteroFlow, an FPGA accelerator programming model that decouples the algorithm specification from optimizations related to orchestrating the placement of data across a customized memory hierarchy. Specifically, we introduce a new primitive named .to(), which provides a unified programming interface for specifying data placement optimizations at different levels of granularity: (1) coarse-grained data placement between host and accelerator, (2) medium-grained kernel-level data placement within an accelerator, and (3) fine-grained data placement within a kernel. We build HeteroFlow on top of the open-source HeteroCL DSL and compilation framework. Experimental results on a set of realistic benchmarks show that, programs written in HeteroFlow can match the performance of extensively optimized manual HLS design with much fewer lines of code.

The ongoing shift of cloud services from monolithic designs to mi- croservices creates high demand for efficient and high performance datacenter networking stacks, optimized for fine-grained work- loads. Commodity networking systems based on software stacks and peripheral NICs introduce high overheads when it comes to delivering small messages. We present Dagger, a hardware acceleration fabric for cloud RPCs based on FPGAs, where the accelerator is closely-coupled with the host processor over a configurable memory interconnect. The three key design principle of Dagger are: (1) offloading the entire RPC stack to an FPGA-based NIC, (2) leveraging memory interconnects instead of PCIe buses as the interface with the host CPU, and (3) making the acceleration fabric reconfigurable, so it can accommodate the diverse needs of microservices. We show that the combination of these principles significantly improves the efficiency and performance of cloud RPC systems while preserving their generality. Dagger achieves 1.3 − 3.8× higher per-core RPC throughput compared to both highly-optimized software stacks, and systems using specialized RDMA adapters. It also scales up to 84 Mrps with 8 threads on 4 CPU cores, while maintaining state-of- the-art µs-scale tail latency. We also demonstrate that large third- party applications, like memcached and MICA KVS, can be easily ported on Dagger with minimal changes to their codebase, bringing their median and tail KVS access latency down to 2.8 − 3.5 us and 5.4 − 7.8 us, respectively. Finally, we show that Dagger is beneficial for multi-tier end-to-end microservices with different threading models by evaluating it using an 8-tier application implementing a flight check-in service.

FPGA-based accelerators are increasingly popular across a broad range of applications, because they offer massive parallelism, high energy efficiency, and great flexibility for customizations. However, difficulties in programming and integrating FPGAs have hindered their widespread adoption. Since the mid 2000s, there has been extensive research and development toward making FPGAs accessible to software-inclined developers, besides hardware specialists. Many programming models and automated synthesis tools, such as high-level synthesis, have been proposed to tackle this grand challenge. In this survey, we describe the progression and future prospects of the ongoing journey in significantly improving the software programmability of FPGAs. We first provide a taxonomy of the essential techniques for building a high-performance FPGA accelerator, which requires customizations of the compute engines, memory hierarchy, and data representations. We then summarize a rich spectrum of work on programming abstractions and optimizing compilers that provide different trade-offs between performance and productivity. Finally, we highlight several additional challenges and opportunities that deserve extra attention by the community to bring FPGA-based computing to the masses.

A primary barrier to rapid hardware specialization with FPGAs stems from weak guarantees of existing CAD tools on achieving design closure. Current methodologies require extensive manual efforts to configure a large set of options across multiple stages of the toolflow, intended to achieve high quality-of-results. Due to the size and complexity of the design space spanned by these options, coupled with the time-consuming evaluation of each design point, exploration for reconfigurable computing has become remarkably challenging. To tackle this challenge, we present a learning-assisted autotuning framework called LAMDA, which accelerates FPGA design closure by utilizing design-specific features extracted from early stages of the design flow to guide the tuning process with significant runtime savings. LAMDA automatically configures logic synthesis, technology mapping, placement, and routing to achieve design closure efficiently. Compared with a state-of-the-art FPGA-targeted autotuning system, LAMDA realizes faster timing closure on various realistic benchmarks using Intel Quartus Pro.

Mainstream FPGA tools contain an extensive set of user-controlled compilation options and internal optimization strategies that significantly impact the design quality. These compilation and optimization parameters create a complex design space that human designers may not be able to effectively explore in a time-efficient manner. In this work we describe DATuner, an open-source extensible distributed autotuning framework for optimizing FPGA designs and design automation tools using an ensemble of search techniques managed by multi-armed bandit algorithms. DATuner is designed for a distributed environment that uses parallel searches to amortize the significant runtime overhead of the CAD tools. DATuner provides convenient interface for extension to user-supplied tools, which enables the end users to apply DATuner to design tools/flows of their interest. We demonstrate the effectiveness and extensibility of DATuner using three case studies, which include clock frequency optimization for FPGA compilation, fixed-point optimization, and autotuning logic synthesis transformations.