Custom Solution to Support a Leading Internet Company’s Multi-Platform AI Business

Industry: Internet, Cloud Service Provider

Customer: Kakao

Product type: GPU-accelerated AI server, general purpose compute server

Introduction

Kakao is a global internet company based in South Korea, best known for its flagship platform KakaoTalk which supports free mobile instant messaging and call features for over 220 million users worldwide. Kakao’s sweeping business model also includes numerous platforms spanning travel, mobile banking, AI/blockchain, games, music, video, maps, and transportation.

Challenge

Throughout Kakao’s expansion, it required new technology and platform development to meet the needs of its multifaceted business, growing scope and user base, massive data volumes and daily maintenance. These factors demand higher requirements on the computing performance, interconnection and scalability of the AI servers supporting the Kakao business. Artificial intelligence became one of its foremost priorities with its products Kakao i (AI platform) and Kakao Mini C (smart speakers). Kakao i’s capabilities include voice, visual and dialogue recognition, recommendation, translation, etc., which Kakao Mini C integrates to provide content recommendations like news, stock exchange, multimedia, and real-time popular searches.

Solution

Artificial intelligence development requires competitive features, model training and a comprehensive technology ecosystem. At the initial establishment of Kakao’s AI Lab, Aivres provided Kakao with a product portfolio solution – the high-density KR4268 inference server at the core, supplemented by KR1280 and KR2280 general purpose servers – and passed the first test verification, thereby helping to kick off Kakao’s AI business model.

1) Product portfolio for various scenarios and a complete ecosystem

In order to satisfy the data layering and data storage demands of the client’s business, Aivres combined several different products: general-purpose KR1280 configured with different hard disks for data layering optimized the computation, operation and management functions of Hadoop cluster data, while the KR2280 model provided flexible configuration and high-density deployment. In addition to supporting multiple types of hard drives for data storage, it also provided audio service support by matching 4 TA GPU cards. When it came to meeting clients’ requirements for power redundancy, the KR4268 was the best choice: 40 units adopting NVLink technology in a single machine can support up to 4 2200W power supplies, 4 PCIe-switches ensure load balancing, and superb scalability and heat dissipation design all contribute to excellent performance. In combining KR1280, KR2280 and KR4268 models, Aivres was able to achieve customer requirements for adapting to various application scenarios like Hadoop, audio services, and AI computing.

2) Complete channel network & self-developed components guarantee agile delivery

Responding to the explosive growth of customer businesses, Aivres established a complete channel network and self-developed components to shorten the procurement cycle and achieve extremely agile delivery. Maintaining regular communication and training with channel partners in the early testing phase enhances their understanding of the systems at hand, and enables distributors to better comprehend customer requirements during the stocking and distribution process. Meanwhile, using components independently developed by Aivres, like SAS and PHY cards, not only shortens the supply cycle, but also reduces the overall cost. By controlling the entire process from production to launch, it took only 25 days from the customer’s order to the final deployment of the solution.

Results and Impact

As the technology continues to iterate and update, Kakao continues to enhance existing AI products Kakao i and Kakao Mini C while developing new AI platforms. By using the KR4268 NVLink technology to build an AI training platform, the cluster performance and linear scalability of the Kakao i AI platform have increased by more than 30% compared with other servers. Kakao Mini C’s human-computer interaction has greatly improved as well. With continued real-time analysis of voice, user habits, and content data, it is able to provide more timely and natural responses to a broader range of user queries, providing a further optimized user experience.

With the computing performance and support provided by Aivres, Kakao is able to continuously analyze data to build user behavior models, depict user portraits, and push better customized content for users. Kakao’s AI products have since made vast improvements in Hadoop’s data operations, audio analysis, and AI calculations, bringing a wider range of AI applications to its business.