The development of Artificial Intelligence (AI) technologies has accelerated in recent years, and the achievements in this field, especially in computer vision, are opening up previously unimaginable possibilities. At the core of this development lies the refinement of neural networks and their support with more powerful computing resources. This article will provide information about NVIDIA's initial steps in the field of AI and the creation of AlexNet, one of the key turning points in this area.
Redaksiya reports that NVIDIA CEO Jensen Huang has explained how the revolution in computer vision began. In 2012, the company was focused on the development of 3D graphics, games, and CUDA cores. At that time, research related to AI technologies was still in its early stages. However, the success of AlexNet prompted NVIDIA to completely change its strategy.
According to Huang, the first machine learning experiments were conducted on ordinary consumer graphics cards, specifically a pair of GTX 580s with 3 GB of memory combined in SLI mode. These graphics cards were primarily designed for gaming and had no special capabilities for accelerating neural networks. However, their parallel architecture proved ideal for such tasks. The development of AlexNet was carried out by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton from the University of Toronto. The researchers were looking for ways to improve image recognition at a time when neural networks were not yet prevalent in their current form.
AlexNet consisted of 8 layers and approximately 60 million parameters and was capable of independent learning through a combination of convolutional and deep neural networks. This model surpassed leading solutions in computer vision by over 70%. Huang noted that AlexNet was optimized for two GTX 580s, and data exchange between the GPUs was only performed when necessary, significantly reducing training time. While discussing this on Joe Rogan's podcast, Huang emphasized that AlexNet worked precisely on this hardware and became a turning point for the entire industry. This event led NVIDIA to invest in machine learning technologies and the creation of devices like DGX, the Volta architecture with its first Tensor Cores, and technologies such as DLSS.