[Technology Toolkit]
No More Short of GPU!

No more short of GPU!

Technology Toolkit 2021 is a technical white paper describing core technologies that are being researched and developed by Samsung SDS R&D Center. We would like to introduce in this paper a total of seven technologies concerning AI, Blockchain, Cloud, and Security with details on their technical definition, key features, differentiating points, and use cases to give our readers some insights into our work.

R&D Cloud for AI

Kubernetes-based GPU clustering technology for AI research/development

Download the report

1. Introduction to Technology

Technology Trends and Background

Since the advent of the commercialization of PCs, demand for computing resources has been growing rapidly day by day. Even the resources that are currently used for basic office work and personal documents are vast compared to the past. Image and video files of course require far more computing resources than that.

So far, the CPU has been used as a major computing resource. With the rapid development in computation capacity, the CPU is increasingly equipped with powerful specifications, thereby satisfying the needs of users. But now that the AI era has arrived, the CPU alone can't afford the computing resources it needs. This is because we now need an overwhelming level of high capacity and high quality work. While the CPU is responsible for processing complex computation, the GPU is responsible for processing simple but large amounts of iterations in parallel. It is completely different from the role of GPU which used to assist the CPU, only helping graphics processing before the AI era.

CPU stands for Central Processing Unit. GPU stands for Graphics Processing Unit. [Figure 1] CPU vs. GPU

When solving a problem with deep learning, the more layers there are, the more variables the computing resource will take into account and iterative calculations to derive more accurate answers. The more layers there are, the more accurate the result will be. However, as the number of layers increases, the required resources increase exponentially. Without the parallel processing of GPU, deep learning is virtually impossible to research with only the existing CPU.

input layer, hidden layer 1, hiden layer 2, hidden layer 3, output layer [Figure 2] Parameters that grow exponentially as layers increase

GPUs required for AI research and development are still expensive. It is a big burden for businesses or schools to buy large amounts of GPUs for AI research. The GPU resources you purchase are not running around the clock with no down time. The GPU remains idle, when there is no user or no action is directed.

R&D Cloud for AI is designed to make the most efficient use of GPU resources. It is a technology that clusters, bundles, and integrates GPUs to utilize and manage. Consolidating GPUs increases overall available capacity. Applying distribution technology to clustered resources allows users who need them to use them effectively as needed when they need them. It also supports automatic environment setting to reduce the time spent in the AI development environment.

Now, users don't have to buy GPUs as much as they need. With R&D Cloud for AI, you can take full advantage of the GPU already in place. It allows you to reduce the time it takes to purchase equipment and set up the environment, and to make AI research and development efficient.


The concept of R&D Cloud for AI is simple. It is a technology that increases the number of available resources by bundling multiple GPUs, and reduces the idle time of assets that occur when operating separately. In order to provide R&D Cloud for AI as a service, clustering, resource distribution, and user convenience and scalability enhancement technology are required. It consolidates computing resources with GPU clustering, and supports users to conveniently and efficiently utilize the resources they need when they want to use the service.

① GPU Clustering

GPU clustering integrates multiple GPU servers with Kubernetes-based technology. As the most fundamental technology of R&D Cloud for AI, the key is not just to tie together multiple resources, but to enable stable resource utilization in a state where multiple resources are integrated.

Before clustering, it is just tie together multiple="multiple" resources. But after clustering, it enables stable resource utilization in a state where multiple="multiple" resources are integrated. [Figure 3] before vs. after GPU clustering

② Job scheduling

Let's say a user starts working with a clustered GPU. If there are few users, there is no problem with working with the GPU. However, when there are many users, the clustered GPU cannot process jobs at once to meet the needs of all users. Depending on the size of the job, the required resources, and the expected time, some jobs are processed immediately, and other jobs need to wait.

Job scheduling is a technology that determines the order in which jobs are allocated and processed by determining the current amount of available resources on the clustered GPU server and identifying the characteristics of incoming jobs. R&D Cloud for AI transforms the basic scheduler provided by Kubernetes to suit the research environment, and additionally develops and applies the required scheduler.

Without job scheduling, the clustered GPU cannot process jobs at once to meet the needs of all users. [Figure 4] Effect of Job Scheduling Application

③ Distributed Computing

As Deep Neural Network (DNN) technology advances, more and more computations are required to be processed. State-of-the-art language processing models, etc., are complex enough to not be able to compute at the right time even with expensive GPUs. DNN performs the task of finding parameters through complex operations. LeNet-5, released in 1998, finds 60,000 parameters, while GPT-3, released in 2020, finds 175 billion parameters. Distributed computing technology is a technology that simultaneously utilizes multiple GPUs to process complex computations that cannot be processed with a single GPU. In fact, using GPT-3 for research requires technology that can simultaneously use 1,000 or more GPUs on multiple machines.

R&D Cloud for AI can help researchers who need to perform complex computations quickly achieve results by introducing distributed learning technologies available in various AI frameworks.

With a single GPU you can only handle 4.7 jobs per day. But with the distributed computing technology, you can handle 35 jobs per day with 8 GPUs or 106 jobs per day with 3x8 GPUs. [Figure 5] Distributed training using multimode GPS at the same time

2. Key Features

R&D Cloud for AI provides various functions to make using clustered GPUs more convenient and effective for users.

Personalized Machine Learning R&D Environment

If you need to carry out a project using a large amount of GPUs, such as in machine learning research, you should of course purchase GPU equipment with the required specifications. However, even if you buy GPU equipment, the environment and pattern are different for each user, so the number of users who can use the purchased equipment is bound to be limited. Even if someone tries to develop new development using the GPU, it will be difficult to configure and set up the necessary environment.

R&D Cloud for AI provides personalized container, for example, one container with TensorFlow, Pandas, and Anaconda, another contains with PyTorch, Numpy, Beautifulsoup, the other with TensorFlow, Pandas, Beautifulsoup. [Figure 6] Examples of personalized containers provided by R&D Cloud for AI

R&D Cloud for AI reduces the time required to set up the environment by providing the user's desired environment as a virtualized container. It is also possible to reuse the environment set by other users by utilizing the characteristics of the clustering environment, or to change and use the required configuration.

AI Optimized Job Scheduling

It would be nice if a clustered GPU could handle all users and jobs all the time, but the reality isn't. Most companies and schools always run out of resources, and even if GPUs are clustered and provided, it is natural that the total amount of clustered resources is less than the total amount of jobs. Scheduling technology is a technology that distributes jobs that need to be queued according to the set policy, in what order and in which node to process.

Bin packing scheduler application more efficient than the default method. It adjusts the order and placement of jobs, reducing the time to complete the entire job execution. [Figure 7] The scheduler provided by Kubernetes vs. the bin packing type scheduler applied to R&D Cloud for AI

R&D Cloud for AI has 3 schedulers specialized for AI research. Each scheduler distributes the most efficient processing according to fairness or job characteristics. R&D Cloud for AI is developing AI Optimized Scheduler based on various schedulers developed so far. When Samsung SDS reaches the target scheduling technology level, R&D Cloud for AI can maximize resource utilization efficiency by adjusting the order and placement of jobs, reducing the time to complete the entire job execution.

Real-time Monitoring

The R&D Cloud for AI can secure and cluster additional GPUs as needed. The service provider must determine whether the currently clustered GPU server is oversupplied, appropriate, or should be additionally equipped. R&D Cloud for AI provides real-time monitoring to check usage and users by server and node. Service providers and administrators can easily see how much assets are being used by using thsese monitoring tools, without the need to read code when aggregating demand needs.

The monitoring function is also useful for users. You can determine when to use R&D Cloud for AI through monitoring, and check which nodes are occupied by which projects in real time. In addition, it shows the core utilization rate of the allocated GPU, so that you can check and improve whether the training task is efficient or if there is any performance degradation due to I/O, etc.

The left shows the average GPU usage rate as a whole, the right shows the average GPU usage rate by each nodes. [Figure 8] Part of the real-time monitoring screen of R&D Cloud for AI

3. Differentiating Points

Samsung SDS's R&D Cloud for AI is meaningful in that it does not utilize the existing Kubernetes-based technologies and various technologies as they are, but has built an optimized environment for AI R&D environment. It provides separate pages for schedulers and monitoring tools to enhance operations. In addition, the AI framework, which AI researchers need the most, is applied, and 10 types of environment assets are provided so that machine learning environments can be easily set up, which is the differentiating point of R&D Cloud for AI.

4. Use Case

Support for Samsung SDS R&D AI R&D Environment

Prior to the establishment of R&D Cloud for AI, Samsung SDS R&D Center had difficulties in purchasing and managing GPUs. This is because we predicted and purchased the GPU specifications and quantity required for each project, but whenever the project ended or the scope changed, who should manage the GPU became unclear. In addition, we even encountered heat and noise problems as researchers worked with the GPU server in their seats.

Since the establishment of the R&D Cloud for AI, we have been able to physically integrate these assets into data centers to store and manage them. The R&D Cloud for AI enables us to start developing right away without having to go through the asset purchase process when launching a project. Since the R&D Cloud for AI is in charge of purchasing and managing assets, researchers can focus on the AI research itself, and thereby reducing the time spent on environment preparation.

After completing the first development in 2020, Samsung SDS R&D Center applied its own pilot service to provide considerably satisfactory convenience to AI R&D and improved functions and performance through the Voice of Customer (VoC) received. In 2021, we will increase the number of equipment to accelerate research and development, and at the same time, we will work on the Computing, Network, and Storage available based on Kubernetes.

5. Closing

R&D Cloud for AI provides a realistic alternative to building a GPU-based research environment where demand is constantly increasing. The various functions provided for AI R&D allow you to make the most of the resources you have when you need to purchase servers, by gradually increasing servers without purchasing huge amounts of resources at once. However, the current R&D Cloud for AI needs to be further improved. In particular, networks are an area of research that requires performance improvement.

The R&D Cloud for AI has great significance in that it has achieved technological results by improving simple and clear points. Starting from the simple concept of ‘integrated use', we continue to expand our technology to find more realistic and effective solutions.

R&D Cloud for AI video screenshot image 5. R&D Cloud for AI
Play Video

▶  The content is proected by law and the copyright belongs to the author.
▶  The content is prohibited to copy or quote without the author's permission.

Inseok SEO
Inseok SEO Cloud

Cloud Research Team at Samsung SDS R&D Center

Inseok Seo is in charge of planning of Samsung SDS Cloud Research Team. He is also an SDS reporter. He explains cloud technologies including R&D Cloud for AI in simple terms.

If you have any inquiries, comments, or ideas for improvement concerning technologies introduced in Technology Toolkit 2021, please contact us at techtoolkit@samsung.com.