SberCloud ML Space gets a major upgrade
11 November 2021
Time to reade: 3 minuets
SberCloud, a provider of cloud technology, presented its upgraded ML Space platform. The announcement was made by David Rafalovsky, Executive Vice President of Sberbank, CTO of Sberbank and Head of Technology, at AI Journey 2021, an international conference.
The most important change to ML Space is that it has become a multi-cluster platform. Users can train their models using the capacities of Christofari and Christofari Neo, Russia's two most powerful supercomputers. They can also use individual V100 or A100 GPU cards for training. This enables agile development and cost management depending on the user's needs and budget.
In double precision, Christofari Neo, the new supercomputer, has an effective performance of nearly 12 petaflops. It has the most powerful A100 GPU cards for up to 80 gigabytes of memory.
Another important upgrade of the platform includes the huge addition to the library of base Docker images for laptops and tasks in ML Space. Now, the customers can create their own Docker images by using the base images and run them without errors in ML Space. All this offers a great benefit to data scientists by reducing their non-productive time.
One other upgrade to ML Space involves AutoML module, which is now available for NLP and Computer Vision tasks. Previously, this automatic model creation module could handle only tabular data, but even then was ranked high in accuracy and performance benchmarks. Now, developers and data scientists can work with new classes of tasks.
David Rafalovsky also presented two new ML Space modules, including AI Services and Pipelines. AI Services is like a mobile app marketplace that allows you to publish and sell the services and products developed by third-party data scientists with an ML component. To do this, you need to upload your model, describe the service of this module of the platform, and all the necessary packaging of the product – such as authorization, authentication, monitoring, etc. All this will be provided by SberCloud free of charge. The module was launched for beta testing, and full monetization of AI services created by third-party experts in machine learning will be available from March 2022.
The Pipelines module allows the experiment cycle to be streamlined and will be in great demand from advanced data scientists. This module saves even more time for a specialist, which can be used for the development of new tasks or data processing.
With its user-friendly modules and services, ML Space enjoys high demand both among the developers of the Sber ecosystem and among data scientists at companies of all sizes or at scientific institutions. In just over a year, it was used by 1,500 individuals from about 250 teams. During that time, they used 2.5 petabytes of data and 300 million GPU minutes.
ML Space makes the development of solutions based on machine learning even more accessible. To celebrate the platform upgrade and launch of Christofari Neo, we increased the amount of the grant for test access to ML Space: 180 thousand rubles will be available to the first 100 corporate customers.
David Rafalovsky, Executive Vice President of Sberbank, CTO of Sberbank and Head of Technology:
«We know that a lot of time is spent by data scientists to prepare tools for ML development. This is why we continue to proactively enrich ML Space with additional modules and features to make it even more user-friendly. We can see how powerful, technology-intensive, and yet affordable it is for developing AI services. Our goal is to make ML development as user-friendly as possible. Let me assure you that we will continue to improve our ML Space and the products based on it».
ML Space with new features and A100 GPU cards will be available to everyone on SberCloud from December 1, 2021. On the SberCloud website, you can get access to ML Space, the ruGPT-3 family of generative models, services and demo stands for developing your own AI services and integrating ready-made SberCloud solutions into your company's products.