VK Cloud launches a service in Kazakhstan with cloud GPU cards for AI and ML scenarios

VK Cloud provides not just GPU power, but a comprehensive infrastructure for data and AI workloads.

Within a single platform, companies can:
— collect and store big data using VK Data Lakehouse
— train and run models on Cloud ML Platform
— use GPUs to accelerate computations.

This enables faster launch of AI projects without the need to assemble infrastructure from multiple providers.

Users will be able to analyze big data, develop neural networks, and train them in days rather than weeks. This creates conditions for the emergence of Kazakhstan’s own high-tech products and services based on AI.

GPU use cases cover a wide range of tasks:
— financial institutions can use accelerators to analyze transactions and detect fraud in real time
— retail and e-commerce companies will gain tools for personalized recommendations and demand forecasting
— in industry and energy, GPUs will assist in modeling complex processes and predicting maintenance needs.

The Cloud GPU service will be particularly useful for universities, research centers, and students working on artificial intelligence projects. Having high-performance GPUs in local data centers allows developing competencies within Kazakhstan while preserving personal data and minimizing infrastructure costs. Startups, in turn, will be able to test hypotheses, train models, and bring ready-made solutions to market faster, enhancing their competitiveness both in Kazakhstan and abroad.

GPUs in the public cloud

Investments in local infrastructure are justified for large enterprises with constant workloads, such as training neural networks or processing big data. However, in most cases, the most cost-effective way to obtain servers with GPUs is to deploy them in the public cloud. A cloud provider offers access to the latest graphics processing units, eliminating the need for regular capital expenditures. This is especially important given that new processors are released every two years, and in the near future, annually.

Another advantage of cloud deployment is flexibility. Cloud resources can be scaled almost instantly. It does not matter whether a company uses a single server or a multi-node cluster with dozens of graphics processing units — cloud capacity increases in minutes.

Features of GPU virtualization in the cloud

There are several approaches to graphics processing unit virtualization, each with its own characteristics and application areas.

vGPU technology. This approach allows one physical GPU to be shared among multiple virtual machines. Each VM is allocated a certain portion of computing power and memory. This technology is used in tasks that do not require full GPU power, or when it is necessary to provide access to graphics accelerators for a large number of users.

GPU passthrough technology. This approach provides direct access to the physical GPU without intermediate abstraction layers. This approach ensures maximum performance and is intended for tasks requiring high graphics performance: from 3D modeling and rendering to cloud gaming.

A managed GPU cluster in Kubernetes can operate in several ways:
— one pod uses one or more GPUs
— resources are distributed among pods via MPS — each process receives dedicated resources
— resources are divided via MIG — each part operates independently.

Cloud GPUs provide scalable resources, integration with other services, and can become part of a comprehensive infrastructure rather than simply replacing on-premises equipment.

INVITRO case study

INVITRO is one of the largest medical laboratory networks in Kazakhstan. Its specialization is laboratory diagnostics and medical services. The company operates in five countries and conducts 260,000 tests daily across 1,930 offices.

To speed up service and reduce errors, INVITRO implemented a Face Detection system based on machine learning. The system automatically recognizes faces and objects by comparing them with a photo database, removing the person from the process and reducing service time.

The pilot project was deployed in two medical offices. INVITRO opted out of local deployment due to the complexity of operating with large volumes of data. The company chose cloud deployment integrated with internal corporate services.

Currently, cameras in medical offices record and send video to the data center. The algorithm sends the most suitable frames for recognition to the computer vision system, which recognizes faces and objects using machine learning from VK Cloud.

The Face Detection system improved customer service quality, reduced time spent searching for and registering clients. As a result, the registration process accelerated, and queues during peak hours disappeared.