Large language model (LLM)-based inference workloads increasingly dominate data center costs and resource utilization. Therefore, understanding the inference workload characteristics on evolving CPU-GPU coupled architectures is crucial for optimization. This paper presents an in-depth analysis of LLM inference behavior on loosely-coupled (PCIe A100/H100) and closely-coupled (GH200) systems. We analyze performance dynamics using fine-grained operator-to-kernel trace analysis, facilitated by our novel profiler SKIP and metrics like Total Kernel Launch and Queuing Time (TKLQT). Results show that closely-coupled (CC) GH200 significantly outperforms loosely-coupled (LC) systems at large batch sizes, achieving 1.9x-2.7x faster prefill latency for Llama 3.2-1B. However, our analysis also reveals that GH200 remains CPU-bound up to 4x larger batch sizes than LC systems. In this extended CPU-bound region, we identify the performance characteristics of the Grace CPU as a key factor contributing to higher inference latency at low batch sizes on GH200. We demonstrate that TKLQT accurately identifies this CPU/GPU-bound transition point. Based on this analysis, we further show that kernel fusion offers significant potential to mitigate GH200's low-batch latency bottleneck by reducing kernel launch overhead. This detailed kernel-level characterization provides critical insights for optimizing diverse CPU-GPU coupling strategies. This work is an initial effort, and we plan to explore other major AI/DL workloads that demand different degrees of CPU-GPU heterogeneous architectures.

Attention Graph Neural Networks (AT-GNNs), such as GAT and Graph Transformer, have demonstrated superior performance compared to other GNNs. However, existing GNN systems struggle to efficiently train AT-GNNs on GPUs due to their intricate computation patterns. The execution of AT-GNN operations without kernel fusion results in heavy data movement and significant kernel launch overhead, while fixed thread scheduling in existing GNN kernel fusion strategies leads to sub-optimal performance, redundant computation and unbalanced workload. To address these challenges, we propose a dynamic kernel fusion framework, DF-GNN, for the AT-GNN family. DF-GNN introduces a dynamic bi-level thread scheduling strategy, enabling flexible adjustments to thread scheduling while retaining the benefits of shared memory within the fused kernel. DF-GNN tailors specific thread scheduling for operations in AT-GNNs and considers the performance bottleneck shift caused by the presence of super nodes. Additionally, DF-GNN is integrated with the PyTorch framework for high programmability. Evaluations across diverse GNN models and multiple datasets reveal that DF-GNN surpasses existing GNN kernel optimization works like cuGraph and dgNN, with speedups up to $7.0\times$ over the state-of-the-art non-fusion DGL sparse library. Moreover, it achieves an average speedup of $2.16\times$ in end-to-end training compared to the popular GNN computing framework DGL.

Reinforcement learning (RL) workloads take a notoriously long time to train due to the large number of samples collected at run-time from simulators. Unfortunately, cluster scale-up approaches remain expensive, and commonly used CPU implementations of simulators induce high overhead when switching back and forth between GPU computations. We explore two optimizations that increase RL data collection efficiency by increasing GPU utilization: (1) GPU vectorization: parallelizing simulation on the GPU for increased hardware parallelism, and (2) simulator kernel fusion: fusing multiple simulation steps to run in a single GPU kernel launch to reduce global memory bandwidth requirements. We find that GPU vectorization can achieve up to $1024\times$ speedup over commonly used CPU simulators. We profile the performance of different implementations and show that for a simple simulator, ML compiler implementations (XLA) of GPU vectorization outperform a DNN framework (PyTorch) by $13.4\times$ by reducing CPU overhead from repeated Python to DL backend API calls. We show that simulator kernel fusion speedups with a simple simulator are $11.3\times$ and increase by up to $1024\times$ as simulator complexity increases in terms of memory bandwidth requirements. We show that the speedups from simulator kernel fusion are orthogonal and combinable with GPU vectorization, leading to a multiplicative speedup.

ABSTRACT Kernel fusion is a popular and effective approach for combining multiple features that characterize different aspects of data. Traditional approaches for Multiple Kernel Learning (MKL) attempt to learn the parameters for combining the kernels through sophisticated optimization procedures. In this paper, we propose an alternative approach that creates dense embeddings for data using the kernel similarities and adopts a deep neural network architecture for fusing the embeddings. In order to improve the effectiveness of this network, we introduce the kernel dropout regularization strategy coupled with the use of an expanded set of composition kernels. Experiment results on a real-world activity recognition dataset show that the proposed architecture is effective in fusing kernels and achieves state-of-the-art performance.