Samsung SDS said on March 23 it launched a GPU-as-a-service (GPU subscription service) based on Nvidia's latest GPU, the B300 (Blackwell Ultra), through its cloud, SCP (Samsung Cloud Platform).

The company said the B300 GPU is equipped with 12-stack HBM3E (high-bandwidth memory), providing 288 GB of memory capacity per GPU and bandwidth of 8 TB per second. It said memory performance improved in AI inference requiring complex computation, with capacity 3.6 times and bandwidth 2.4 times higher than the H100.

Data bottlenecks that degrade overall performance when running large language models (LLMs) due to memory data transfer speeds being slower than GPU compute processing speeds have been dramatically improved. Samsung SDS has provided GPUaaS based on the A100 in 2021 and the H100 in 2023, supporting GPUs to be used as AI-dedicated infrastructure across cloud-based infrastructure buildout, operations and customer services.

Customers adopting the SCP B300 GPUaaS can efficiently process large AI models through high-capacity memory, helping minimise latency for high-performance AI services such as AI agents and image, video and code generation and analysis, the company said.

Samsung SDS also plans to launch in the third quarter this year a serverless inference service in which users pay only for the amount of tokens used, without separate infrastructure usage fees when applying AI models. It also plans to launch an AI training service that automatically and immediately performs distributed training when developers input code and data.

Lee Ho-jun (이호준), head of Samsung SDS' Cloud Service Business Division and a vice president, said, "Based on GPU efficiency capabilities such as resource optimisation and energy savings that SCP has, we will provide South Korea's first B300 GPU service to customers seeking to apply AI to their work, including large companies, mid-sized and small businesses, and the public sector, and actively support their AX transformation."