Application of Generative Artificial Intelligence from Neural Networks in Data Centers

In the era of Artificial Intelligence and specifically in the sub-category of Generative Artificial Intelligence (GenIA), the trend in the context of Data Centers is to be able to integrate traditional network architectures that are still managed as Ethernet networks with new ones. GenIA-based network structures that use low latency protocols such as InfiniBand and that allow parallel processing. In this work, a novel network architecture is presented that integrates the relevant aspects of Ethernet networks and the new GenIA Back End networks. We also deal with aspects related to protocols, bandwidth, design of a GenIA network architecture, components of GenIA networks and cooling systems. The main objective of this article is to analyze, through the Radial Basic Function (RBF) and Multilayer Perceptron (MLP) algorithms, which are part of the scope of Neural Networks (NN), how the dependent variable or metric GPU Performance behaves, since it is one of the critical aspects in the middle of modern data center architectures. To achieve this, a series of independent variables are defined such as GPU Memory Access, Memory Bandwidth, Throughput, Power Consumption, Temperature, and Clock Speed. With the RBF algorithm, these sub-metrics are analyzed taking into consideration the number of units that may be in each of the hidden layers and the level of importance of the variable, while, with the MLP, the number of units established both in hidden layer 1 and 2, respectively, and the importance level and normalized importance to analyze the behavior of these sub-metrics. The latter is presented through case studies and helps to identify which metrics are more relevant than others in terms of operation, performance, and efficiency of GPU performance metrics.