Large Language Models (LLMs) have demonstrated remarkable performance in natural language processing and are now widely adopted in diverse applications. However, their significant computation and memory costs severely limit their acceleration. In particular, the self-attention mechanism is a significant bottleneck, as it cannot exploit batch parallelism across prompts, and its memory traffic grows quadratically with sequence length. In this paper, we propose Scrooge, a novel hardware accelerator framework that leverages an attention early termination mechanism, designed to address the inefficiency of self-attention. The self-attention mechanism does not assign equal importance to all tokens. Instead, semantically important tokens consistently receive higher attention scores. Consequently, preserving sufficient attention for a subset of important tokens is often enough to maintain model accuracy, even without computing attention for all tokens. Our key insight is that once sufficient attention has been accumulated, further computation with the remaining tokens only increases complexity without improving accuracy. Scrooge leverages this insight to approximate the attention of the remaining tokens and terminates the attention computation dynamically once it has gathered sufficient attention. With this method, Scrooge reduces both latency and memory traffic while maintaining accuracy. Experimental results show that Scrooge achieves a 1.7× speedup and a 0.47× reduction in memory traffic with negligible accuracy loss.

LLM inference often exceeds on-chip memory capacity, causing frequent external memory access. Quantization reduces memory cost but loses accuracy due to outliers. Prior mixed-precision accelerators address this issue with encoding schemes, but often result in accuracy degradation for LLMs and pipeline stalls. We present Lupin, an algorithm-architecture co-design with Outlier-First Encoding, which stores outliers in high precision by reallocating less critical normal values. This preserves maximal outlier representation and enables stall-free execution with low-precision MAC units. Experiments show that Lupin maintains accuracy while achieving a 2.02× speedup.

Online data scraping has taken on new dimensions in recent years, as traditional scrapers have been joined by new AI-specific bots. To counteract unwanted scraping, many sites use tools like the Robots Exclusion Protocol (REP), which places a robots.txt file at the site root to dictate scraper behavior. Yet, the efficacy of the REP is not well-understood. Anecdotal evidence suggests some bots comply poorly with it, but no rigorous study exists to support (or refute) this claim. To understand the merits and limits of the REP, we conduct the first large-scale study of web scraper compliance with robots.txt directives using anonymized web logs from our institution. We analyze the behavior of 130 self-declared bots (and many anonymous ones) over 40 days, using a series of controlled robots.txt experiments. We find that bots are less likely to comply with stricter robots.txt directives, and that certain categories of bots, including AI search crawlers, rarely check robots.txt at all. Our findings suggest that relying on robots.txt to prevent unwanted scraping is risky and highlight the need for alternatives.

As the information technology plays an important role in the smart factories, Ethernet-based industrial network has rapidly replaced the traditional field buses. To maintain this critical network secure, it is important to develop the realtime network intrusion detection system (NIDS). The widely used NIDS was developed for the general Internet environment where the average throughput to protect attacks from the large number of unknown network nodes is more important than the real-time detection capability. However, in the critical industrial network, the real-time protection is more important than the average throughput. In this paper, a FPGA-based abnormal Ethernet packet detector is proposed. Since it is designed for the closed industry network, packet detection is based on the whitelist that consists of the allowed network address and protocol numbers. The prototype system has been implemented using the Xilinx Zynq-7030 SoC running at 250MHz. The network header of the Ethernet packet is compared to the 256 whitelist ruleset within 0.032μsec, which means that the malicious packets from the abnormal network nodes are filtered out even before the whole packets arrives. This real-time packet filtering feature is useful in protecting highly secure network systems like the critical industrial control systems.