Performance Analysis of AI Models Across GPUs and SDKs: A Benchmarking Approach

Main Article Content

Santosh Appachu Devanira Poovaiah
Sai Sharath Yadagiri

Abstract

The selection of hardware and software platforms has a significant impact on how well AI models are trained. Despite the widespread usage of several GPU architectures and deep learning frameworks like PyTorch, TensorFlow, and JAX, there is a dearth of thorough testing data to help developers choose the best GPU–SDK pairings. This paper presents a systematic empirical study evaluating the performance of four representative AI models—ResNet-18, BERT-Base, GraphSAGE, and GPT-2 Small—across three prominent SDKs and three commonly available GPUs: NVIDIA A100, V100, and T4. Key metrics including training time, memory consumption, and throughput are analyzed to reveal the impact of hardware-software pairings on training efficiency. The results demonstrate significant performance variations based on model architectures and GPU–SDK choices, providing actionable insights for AI practitioners to optimize resource utilization and reduce training costs. This work fills an important gap by offering data-driven guidance for selecting hardware and software configurations in AI model training [1] [2].

Article Details

Section
Articles