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.

The existing neural architecture search (NAS) methods usually restrict the search space to the pre-defined types of block for a fixed macro-architecture. However, this strategy will limit the search space and affect architecture flexibility if block proposal search (BPS) is not considered for NAS. As a result, block structure search is the bottleneck in many previous NAS works. In this work, we propose a new evolutionary algorithm referred to as latency EvoNAS (LEvoNAS) for block structure search, and also incorporate it to the NAS framework by developing a novel two-stage framework referred to as Block Proposal NAS (BP-NAS). Comprehensive experimental results on two computer vision tasks demonstrate the superiority of our newly proposed approach over the state-of-the-art lightweight methods. For the classification task on the ImageNet dataset, our BPN-A is better than 1.0-MobileNetV2 with similar latency, and our BPN-B saves 23.7% latency when compared with 1.4-MobileNetV2 with higher top-1 accuracy. Furthermore, for the object detection task on the COCO dataset, our method achieves significant performance improvement than MobileNetV2, which demonstrates the generalization capability of our newly proposed framework.

Example weighting algorithm is an effective solution to the training bias problem, however, most previous typical methods are usually limited to human knowledge and require laborious tuning of hyperparameters. In this paper, we propose a novel example weighting framework called Learning to Auto Weight (LAW). The proposed framework finds step-dependent weighting policies adaptively, and can be jointly trained with target networks without any assumptions or prior knowledge about the dataset. It consists of three key components: Stage-based Searching Strategy (3SM) is adopted to shrink the huge searching space in a complete training process; Duplicate Network Reward (DNR) gives more accurate supervision by removing randomness during the searching process; Full Data Update (FDU) further improves the updating efficiency. Experimental results demonstrate the superiority of weighting policy explored by LAW over standard training pipeline. Compared with baselines, LAW can find a better weighting schedule which achieves much more superior accuracy on both biased CIFAR and ImageNet.

Most teacher-student frameworks based on knowledge distillation (KD) depend on a strong congruent constraint on instance level. However, they usually ignore the correlation between multiple instances, which is also valuable for knowledge transfer. In this work, we propose a new framework named correlation congruence for knowledge distillation (CCKD), which transfers not only the instance-level information but also the correlation between instances. Furthermore, a generalized kernel method based on Taylor series expansion is proposed to better capture the correlation between instances. Empirical experiments and ablation studies on image classification tasks (including CIFAR-100, ImageNet-1K) and metric learning tasks (including ReID and Face Recognition) show that the proposed CCKD substantially outperforms the original KD and other SOTA KD-based methods. The CCKD can be easily deployed in the majority of the teacher-student framework such as KD and hint-based learning methods.

Medical ultrasound imaging is a typical clinical diagnostic method which has been widely applied to heart movement monitoring, fetal development status tracking, and liver interventional surgery, etc. Among various medical ultrasound imaging algorithms, delay-and-sum algorithm is the most common algorithm which can be found in almost every medical ultrasonic scanner. The execution of delay-and-sum beamforming algorithm is not time-consuming typically. However, when constructing a medical imaging cloud service, the computing speed of the delay-and-sum algorithm on the cloud server has to be much faster than the computing speed of the algorithm implementation on regular ultrasonic scanner processor, so as to overcome the cloud data transfer speed constraint on the real-time imaging requirement for medical ultrasound. As such, this paper demonstrated the implementation of the parallel delay-and-sum beamforming algorithm on the FPGA with coherent accelerator processor interface. Furthermore, the performance evaluation results were also demonstrated in the paper.