Samsung SDS said on the 23rd that it launched GPUaaS (GPU subscription service) based on Nvidia's latest graphics processing unit (GPU), "B300 (Blackwell Ultra)," through its "Samsung Cloud Platform (SCP)," the first such service in Korea.
The service launch is a strategy to meet the surging demand for high-performance computing as corporations move beyond developing artificial intelligence (AI) models and enter the "AI inference" phase of applying them to real services.
The B300 GPU is equipped with 12-stack HBM3E (high-bandwidth memory), providing 288GB (gigabytes) of memory capacity and 8TB (terabytes) per second of bandwidth per GPU. Based on this, in AI inference that requires complex computation, memory performance has improved compared with the H100, with capacity up to 3.6 times and bandwidth up to 2.4 times.
Accordingly, the company said it has dramatically improved the data bottleneck that degrades overall performance due to slower memory data transfer speeds compared with the GPU's fast compute speed when running large language models (LLMs).
Samsung SDS proactively introduced GPUaaS based on the A100 in 2021 and the H100 in 2023, leading the formation of a GPUaaS ecosystem so GPUs can be used as AI-dedicated infrastructure for cloud-based infrastructure buildout, operations, and customer service.
A company official said, "Customers adopting 'SCP B300 GPUaaS' can efficiently handle large AI models through high-capacity memory, minimizing latency for AI services that require high performance, such as AI agents and image, video, and code generation and analysis."
They also noted that using a subscription model that lets you use what you need and pay accordingly can reduce initial investment risk and streamline expense. Even when GPUs are hard to procure, customers can immediately adopt Nvidia's latest architecture for their work through SCP, and, the company emphasized, handle sensitive corporate data in a secure cloud environment combined with Samsung SDS's security capabilities.
Samsung SDS plans in the third quarter to launch a "serverless inference service," which charges expense only for the number of tokens used with no separate infrastructure usage fee when applying an AI model, and an "AI training service," which automatically and instantly distributes AI training when developers input code and data.
Lee Ho-jun, head of the Cloud Service Business Unit (executive vice president) at Samsung SDS, said, "Based on SCP's GPU optimization capabilities, such as resource optimization and energy savings, we will actively support AX (AI transformation) by providing the first B300 GPU service in Korea to customers seeking to apply AI to their work, including large corporations, mid-sized and small businesses, and the public sector."