一句话：首个利用稀疏表示的自动机处理加速器 Hopps，混合压缩与未压缩单元，等面积下显著超越 SOTA。
自动机处理（AP）是数据分析与科学计算的关键内核，但现有加速器多以位集无压缩存储自动机图，依赖交叉开关获得高并行，对稀疏图效率低。Hopps 是首个利用稀疏数据表示的 AP 加速器：它混合两类处理单元，一类无压缩、吞吐高但空间效率差，另一类采用压缩稀疏表示、空间高效但吞吐较低且波动。为发挥优势，论文提出自动机映射算法，将大部分工作分给高吞吐单元，同时把大量状态留在空间高效单元，并放宽交叉开关设计约束，允许使用大量较小交叉开关。在 AutomataZoo 上，等面积下 Hopps 相对 Impala 和 SpAP 的几何平均加速分别为 2.5 倍和 2.2 倍。
Automata processing (AP) is a key kernel in data analytics and scientific computing. AP workloads process a stream of symbols with many automata (FSMs) in parallel, e.g., pattern-matching network traffic against many malicious strings. The need for high-performance AP has sparked the design of specialized accelerators. But prior AP accelerators are inefficient: AP workloads have substantial sparsity, but accelerators exploit no or limited sparsity. Specifically, each AP workload can be expressed as the concurrent traversal of all automata, which are encoded as graphs. But state-of-the-art accelerators store these graphs uncompressed, using bitsets. This allows the use of specialized memory crossbars that provide high parallelism and efficiency when graphs are dense. But many graphs are highly sparse, making crossbar-based accelerators inefficient. We present Hopps, the first automata processing accelerator that exploits sparse data representations. Hopps combines two types of processing units: one represents data uncompressed, which achieves high throughput but is space-inefficient, while the other uses a compressed-sparse representation, which achieves high space efficiency but lower and more variable throughput. To use Hopps well, we present a novel automata mapping algorithm that maps most work to high-throughput units, while keeping a large fraction of state in space-efficient units. Hopps's hybrid design relaxes several constraints in crossbar-based designs, allowing for more efficient high-throughput units (e.g., by using a large number of smaller crossbars). Thus, by making the uncommon case cheap, Hopps makes the common case even faster. We evaluate Hopps on AutomataZoo benchmarks. Hopps outperforms prior state-of-the-art accelerators Impala and SpAP by gmean 2.5x and 2.2x when using equal area.