The rapid progress in Deep Learning (DL) and Large Language Models (LLMs) has exponentially increased demands of computational power and bandwidth. This, combined with the high costs of faster computing chips and interconnects, has significantly inflated High Performance Computing (HPC) construction costs. To address these challenges, we introduce the Fire-Flyer AI-HPC architecture, a synergistic hardware-software co-design framework and its best practices. For DL training, we deployed the Fire-Flyer 2 with 10,000 PCIe A100 GPUs, achieved performance approximating the DGX-A100 while reducing costs by half and energy consumption by $40 \%$. We specifically engineered HFReduce to accelerate allreduce communication and implemented numerous measures to keep our Computation-Storage Integrated Network congestion-free. Through our software stack, including HaiScale, 3FS, and HAI-Platform, we achieved substantial scalability by overlapping computation and communication. Our system-oriented experience from DL training provides valuable insights to drive future advancements in AI-HPC.

Graph convolution networks (GCNs) have achieved impressive results for few-shot hyperspectral image classification. However, current methods focus on migrating labels from support patches to query patches through graph convolution. Less attention has been paid to the patch features generated by graph convolution. This leads to that the methods can’t perceive the abnormal situation that the labels of two patches from different classes are correctly predicted but their features are also similar with each other. Moreover, current methods usually adopt multiple layers of graph convolution for label migration and have not distinguished the impact of different layers. This may result in that important information are submerged in the iterative convolution process. To address these issues, this paper proposes a Contrastive Graph Convolution Network with Skip Connection for few-shot hyperspectral image classification (CGCN-SC). By following the supervised contrastive learning, the contrastive loss modules are proposed to constrain the relationship between the generated features of the patches. It is expected that while migrating labels from labeled patches to unlabeled ones, the generated features of patches from the same classes are close with each other and the features from different classes are far away. Moreover, the skip connection is proposed to be incorporated into the graph convolution network to allow high-level abstract information to be combined with low-level features that may be lost during graph convolution. Through extensive experiments, the proposed method has shown better performance when compared with a set of related methods. All the codes are available at github https://github.com/dongao11/CGCN.

Compressed sensing (CS) techniques can reduce the sampling rates required in synthetic aperture radar (SAR). However, it is difficult to use the restricted isometry property to theoretically analyze the performance. Therefore, in this letter, information theory is applied to set necessary bounds on sampling ratios in CS-based SAR imaging. The system is viewed as a multi-input/multi-output (MIMO) channel, with information capacity quantified for a given measurement matrix and signal-to-noise ratio (SNR). According to the source-channel coding theorem, the lower bound of the sampling ratios is derived in terms of sparsity ratio, SNR, bandwidth, and radar pulse duration. Simulation studies are performed to test and analyze the information-theoretical bounds.

With the rapid growth of web data, people sometime need semantic similar information in order to obtain a clear outline of their interests, so recommendation is needed to provide relevant information to users' queries. In this paper, we propose a method to recommend semantic similar movies and stars to users' queries, styles and stories. The system measures the similarities between movies according to genre and style features extracted from YAGO and IMDB. Experimental results show that the recommendations meet users' interests.