By leveraging KAYTUS KR4268V2 alongside MotusAI, assisted a well-known Korean gaming company to build a large language model

KAYTUS KR4268V2 has powerful computing power, combined with MotusAI, can shorten model training time by over 700 hours and accelerate the rapid deployment of large language models.

Customer Introduction

This prominent game manufacturer is a Korean gaming giant with its headquarters based in Seoul. It has extended its subsidiaries and joint ventures in many countries namely the United States, Europe, and Asia. The main development direction of this company is the large-scale multiplayer online role-playing games, and its operating games include multiple world-class games. Based on statistics, this manufacturer ranks in the top 30 of global mobile game rankings.

Project Background

Currently, global companies are extensively exploring the optimal utilization of generative AI  to generate profits. The potential of AI to enhance communication and foster creativity across diverse domains.  In response to current AI trends, the gaming company has been conducting research in AI and NLP fields and plans to develop a large language model for the Korean people. To address this evolving demand, a robust server infrastructure is urgently needed to support its large and complex language model.

Business Challenges

1. A high-performance and high-scalability IT infrastructure

(1) Large language models are built based on the deep learning approach,  and are necessary to process massive amounts of data, which usually require complex data cleaning, sensitive data analysis, semantic understanding, and so forth. High-performance IT infrastructure can provide faster processing capacity, as a result, the data processing time is significantly minimized and enhance the work efficiency.

(2) Large language models require storing large amounts of data and need to perform constant read and write operations on the data. Highly scalable IT infrastructure can provide greater storage capacity and better scalability to meet the continuously growing demand for data storage.

(3) With the rapid business development and surge of data volume, the adaptability of large language models becomes crucial in the context of their flexible processing capability and storage capacity. A highly scalable IT infrastructure can provide dynamic scaling solutions to meet the rigorous changing requirements.

2. Computing resource utilization

 The computational requirements for large language models are typically enormous, as these models generally contain hundreds of millions to billions of parameters. Clients have conducted numerous research projects in the fields of text generation, text-to-image generation, digital human creation, and conversational agent development. In order to mitigate risks and ensure business security, these projects must undergo extensive development and continuous trial and error. Correspondingly, many AI models require repeated adjustments and optimizations, undergoing tens or even hundreds of iterations of training, before achieving the required accuracy and safety for business pilot testing. The high demand for accuracy and security has led to a rapid increase in computational resource consumption, resulting leads to resource congestion and training queuing. While expanding the number of AI servers and augmenting the overall computational resources presents a viable solution, it concurrently amplifies the challenge of managing the additional resources, significantly elevating the complexity of resource management.

Solution introduction

The KR4268V2 AI server series, fabricated with Intel configuration, is a new generation of AI servers designed and developed by KAYTUS for complex application scenarios such as deep learning, met averse, AIGC, and AI + Science. It features diversified computing power and an elastic architecture expansion, making it flexible and applicable to multiple scenarios. The KR4268V2  series, with Intel configuration, servers have undergone full-chain performance optimization, providing a maximum of 13 PCIe 5.0 expansion slots, allowing flexible allocation and topology switching between CPUs and GPUs based on application scenarios. It supports a maximum of two Intel fourth-generation Xeon Scalable processors, with a maximum of 128 processor cores, 8TB system memory, and 300TB local high-speed storage, breaking the bottleneck of memory and storage and providing strong support for AI training and inference. By integrating the MotusAI development and management platform, KAYTUS AI servers can seamlessly connect with various storage systems through the built-in data management module, providing a visual interface for centralized data management. Developers can quickly upload data through the data management interface, and access data in real-time through directory mounting for data preprocessing, model development, and model debugging. After accomplishing the training, users can save, upload, and share model files as required. With the data caching acceleration mechanism, MotusAI can improve model training efficiency and shorten the model development cycle.

Customer Benefits

Leveraging KAYTUS' state-of-the-art, high-performance, and highly scalable hardware, in conjunction with MotusAI, which effectively addresses customers' large language model's computational resource demand and provides robust computing capacity. As a result, the proposed solution significantly shortens the model training time by more than 700 hours and improves training efficiency. Through the efficient resource sharing and optimized scheduling features of MotusAI, the system ensures the better utilization of computing, storage, and network resources. The deployed solution has improved system automation by more than 50%, while meeting customers' rapid launch requirements for Korean big data models.