AI System in Russia Aims to Make Neural Network Development Cheaper
A new system from the Artificial Intelligence Research Institute (AIRI) automatically generates and optimizes GPU code used in training neural networks.
Researchers at the Artificial Intelligence Research Institute AIRI have released an AI-based system that automatically generates and optimizes GPU kernels. These specialized pieces of code run on graphics accelerators and are widely used when training and operating neural networks.
According to the developers, the technology can speed up the development of machine learning models while reducing spending on computing resources. As competition in the artificial intelligence sector intensifies, companies are increasingly trying to make model training more efficient and less expensive.
Optimizing Computation
Modern AI systems rely heavily on graphics accelerators. They run specialized programs known as GPU kernels, which consist of large numbers of mathematical operations. Their performance depends on the structure of the code and on the architecture of the specific graphics card.
Universal libraries rarely deliver maximum performance, so developers often write GPU kernels manually. This process requires time and the involvement of highly skilled specialists.
The system developed at AIRI automates this task. It generates GPU kernel code on its own and optimizes it for the architecture of graphics accelerators. The new tool can accelerate neural network development and reduce the cost of building AI models.
Developers say the technology can also be adapted for other types of GPU code. This would require modifying the system’s source code, which could broaden the tool’s range of applications for other computing tasks.
