Sber Develops ML Space, One of World's Most Affordable Cloud Solutions for AI Model Training

  • Artificial Intelligence , Machine Learning
  • 04.12.2020 07:36 pm

Sber ecosystem resident SberCloud has developed Machine Learning Space (ML Space),a cloud platform forAI model training, David Rafalovsky, Executive Vice President, CTO of Sberbank Group, Head of Technology, said at the AI Journey conference.

“We at Sber believe that instead of being 30%-focused on training models, Data Scientists should devote 99%of their time to thisand be able to work from anywhere in the world using the cloud. That’s why we have created this solution. Today I am happy to present Machine Learning Space by SberCloud,” said David Rafalovsky.

First, he said, DS professionals will enjoy a dramatic acceleration of their experiments. “Thanks to the Christofari, resource-intensive models will take mere hours to train. Not weeks or months. To that end, we have created a very convenient module, called Environments, on the platform. It is used to launch training using the Jupyter Notebook or JupyterLab you are used to. As a learning environment, you can use any of your Docker images. The module contains all the utilities you’d need to monitor resource consumption, models, and experiments. I’m talking about Grafana, TensorBoard, and MLflow. Thus you can easily monitor the progress of model training per epoch while also comparing the results with other experiments quickly,” David Rafalovsky stated.

He said that the Christofari supercomputer, which does all the computing, is licensed to work with personal data, so models can be trained on such datasets.

Second, he said, ML Space allows teams of any size to work together, regardless of their location. “This is especially important today when many of us telework while facing more tasks. Our platform is unique because of the versioning of datasets and files.And in March, version control of Docker images, scripts, and models will be ready. All your work is now structured and transparent even if dozens of Data Scientists work with datasets or code. The platform’s Data Catalog module is responsible for this letting you set up access rights to data and manage them,” Sberbank CTO added.

He also said that a collaborative environment is scheduled to be unlocked in March to let stakeholders work with models and ML artifacts from external sources. “To this end, we will launch two more modules, Model Registry and ML DataHub. Model Registry is a centralized repository of models with tools to manage their entire lifecycle, letting users share models and place them in the AutoDeploy module with a few clicks. Datasets, models, Docker containers, and other artifacts for popular ML tasks, likeOCR or NER,will be available in ML DataHub. You’ll no longer need to waste time searching and downloading models or containers from external sources, or check them for viruses scrupulously, as ML DataHub will contain secure dataonly,” said David Rafalovsky.

According to Mr. Rafalovsky, ML Space will be available on SberCloud starting from December 12. “As for the price, it will be one of the lowest in the world. Moreover, all ML Space modules will use the Pay-As-You-Go pricing model. For example, the cost of data storage will start from 1 ruble 20 kopecks a month per gigabyte, CPU preprocessing and training of models will be priced from 12 kopecks per minute, and the price for GPU model inference will start from 5 kopecks per second. It more than affordable, it is incredibly cheap,” David Rafalovsky said in conclusion.

ML Space is already being used by Sber, companies of its ecosystem, and some external partners.

Related News