Shouxu Lin

Shouxu Lin, Zhiyuan Guo, Jiaxin Lin

LLM inference is constrained by GPU memory capacity and bandwidth. Tiered memory architectures mitigate this by allowing the GPU to offload memory to the remote tier. However, existing memory offloading frameworks rely on prefetching data into local GPU HBM. This approach underutilizes system resources by introducing HBM contention, squandering memory capacity, and creating pipeline bubbles. We show that enabling direct GPU access to remote memory significantly outperforms prefetching, achieving optimal aggregate system bandwidth. We propose DAK, an end-to-end direct-access memory offloading framework that repurposes the Tensor Memory Accelerator (TMA) to asynchronously fetch offloaded weights and KV caches directly from remote memory into GPU shared memory (SMEM). To maximize remote access performance, DAK introduces a greedy algorithm to determine optimal per-operation offloading ratios, alongside active congestion control and TMA multicast to eliminate interconnect bottlenecks and read amplification. Evaluations across diverse architectures show that DAK achieves near-optimal bandwidth aggregation, with up to 3$\times$ performance gains on NVLink-C2C and 1.8$\times$ on PCIe systems compared to state-of-the-art memory offloading baselines.

Yang Liu, Tom Gedeon, Sabrina Caldwell et al.

Emotion recognition based on physiological signals is a hot topic and has a wide range of applications, like safe driving, health care and creating a secure society. This paper introduces a physiological dataset PAFEW, which is obtained using movie clips from the Acted Facial Expressions in the Wild (AFEW) dataset as stimuli. To establish a baseline, we use the electrodermal activity (EDA) signals in this dataset and extract 6 features from each signal series corresponding to each movie clip to recognize 7 emotions, i.e., Anger, Disgust, Fear, Happy, Surprise, Sad and Neutral. Overall, 24 observers participated in our collection of the training set, including 19 observers who participated in only one session watching 80 videos from 7 classes and 5 observers who participated multiple times and watched all the videos. All videos were presented in an order balanced fashion. Leave-one-observer-out was employed in this classification task. We report the classification accuracy of our baseline, a three-layer network, on this initial training set while training with signals from all participants, only single participants and only multiple participants. We also investigate the recognition accuracy of grouping the dataset by arousal or valence, which achieves 68.66% and 72.72% separately. Finally, we provide a two-step network. The first step is to classify the features into high/low arousal or positive/negative valence by a network. Then the arousal/valence middle output of the first step is concatenated with feature sets as input of the second step for emotion recognition. We found that adding arousal or valence information can help to improve the classification accuracy. In addition, the information of positive/negative valence boosts the classification accuracy to a higher degree on this dataset.