GPU Cloud

The GPU Cloud-based training of AI models is an uncontested approach to solving machine learning tasks. Here, we show how a GPU Cloud helps optimize and accelerate the process of improving the artificial intelligence models for object recognition, which is related to the processing of large data massive.

Project with artificial intelligence

A company, which we can’t mention under the NDA, had to train AI models to recognize objects based on a vast pool of photos. The task was to adjust the process of AI training as quickly and efficiently as possible and to ensure their accurate work during the real execution.

How to train an AI model

In general, the process of AI models training for object recognition is based on six main stages:

• Data collection and preparation. To educate a model, a large set of marked up objects is required for it to distinguish. Usually, this data consists of thousands of images, each with annotation, so that it includes information on the object’s surroundings and its class (e.g., “a car”, “a human” etc.”).
• Data augmentation. To increase the model’s accuracy, data is often augmented. It means that the image is shifted: rotated, mirrored, with changed brightness and contrast in order to create new image examples. It allows the model to better work with different object variations in practice.
• Model training. The model is trained while going through the images in the training dataset, adjusting its parameters with each iteration.
• Auditing and testing. After the training, the model is given a test database to assess its accuracy and the ability to recognize objects. If the results are unsatisfactory, it is worked upon and new data is added to improve results.
• Deployment. When the model is ready, it is deployed in real conditions, where it processes new images and recognizes objects based on what it learned.

With limited capacities of ordinary CPU processors, such a learning process takes a lot, especially considering the need for manyfold iterations for testing and results improvement. The use of cloud solutions based on NVIDIA graphic processors allows for a significant boost in development and fast deployment in the real life. It is the optimal solution for companies, for which time to market is critical.

Search of a technological partner

Having considered everything mentioned above, the customer had to train on the basis of graphic processors, that is GPUs. For that reason, this company rented out a GPU Cloud from GigaCloud, based on the NVIDIA A40 processors.

Contrary to CPU (central processing unit), which is used for basic operations and computes them one by one, GPU performs computations in parallel, in several streams. Thanks to this, the processes requiring much load are completed faster.

“It was important for us to train models on resources provided by the Ukrainian cloud provider. It guarantees safety and data privacy conformity,” says the company’s CTO. “GigaCloud is one of the few national providers offering a cloud with GPU, and we knew that it was built on the basis of a data center compliant with the TIER III level. So we had no doubts in cooperating with them”.

After moving everything necessary for AI models training into the cloud, including the image database, the customer’s team managed to ca project in just a week: train AI models, test them out, find bugs and finalize everything.

Cloud with GPU is a must-have for AI tasks

Thanks to the GPU Cloud, the client could speed up the machine learning of models 50 times faster than with ordinary CPUs. Moreover, the cloud provided a series of other benefits.

• Fast scale-up. There is a possibility to expand computing capacities by adding resources in a few minutes.
• No need for maintenance. As there is no need to buy, install and service the equipment, the project team can focus on educational tasks.
• Access from any device. Ensures the convenience while working with resources, regardless of the device location or type.

Training, testing, and improving models occur cyclically, enabling the client’s team to respond quickly to changes and continuously optimize object recognition accuracy. After testing the models and deploying them in real-world conditions, the team plans to convert the process into a production line.

“Thanks to GigaCloud’s support, the team was able to focus on developing and refining AI models without wasting resources on purchasing and maintaining physical infrastructure. This allowed for the more efficient use of financial resources while maintaining high performance and flexible scalability,” said the company’s CTO.