KT Cloud Commercializes 'Cloud-Based NPU Infrastructure'

The AI semiconductor infrastructure for research and development of super-large artificial intelligence (AI) can now be easily accessed with just a few clicks through an online-only portal.

KT Cloud announced on the 27th that it has commercialized a cloud-based neural processing unit (NPU) infrastructure service applying 'Atom' from the domestic fabless semiconductor design company Rebellions, and started the service on May 30.

NPU infrastructure service of kt cloud applied with Rebellion's Atom chip
[Photo by KT Cloud]

NPUs are AI semiconductors optimized for AI applications. Compared to graphics processing units (GPUs) of the same class, they offer faster computation speeds and lower power consumption, bringing innovation to AI research and development time and costs. KT Cloud, in collaboration with Rebellions, has commercialized a cloud-based NPU infrastructure equipped with the high-performance NPU ‘Atom,’ which has been recognized globally. KT Cloud integrated the NPU into its platform for cloud-based use, implemented resource sharing pools and auto-provisioning, and turned it into a service. Companies can easily and conveniently perform AI training and inference based on NPU through a dedicated portal.

The cloud-based NPU infrastructure allows customers to create resources, utilize computation sessions, and manage and monitor them directly with just a few clicks, offering higher usability compared to on-premise infrastructure. Additionally, it provides an environment where the high-performance, low-power characteristics of the NPU can be used as needed, bringing innovation in research time and costs for AI companies.

KT Cloud will first provide the NPU infrastructure to companies participating in the ‘AI Voucher Support Project’ and the ‘High-Performance Computing Support Project,’ which are part of the ‘K-Cloud Project’ aimed at supporting early market demand for AI semiconductors and providing computing resources based on domestic AI semiconductors to small and venture companies. The service will be expanded to corporate customers in the second half of this year.

The ‘Atom’ installed in this service is the second AI semiconductor introduced by Rebellions, following ‘Ion,’ an AI semiconductor specialized for financial transactions. It already possesses performance at the level of second-generation NPUs being developed by other domestic fabless companies. Atom is the only domestic NPU that supports transformer language models (machine translation models) and floating-point operations. According to the results of ‘MLPerf,’ an AI semiconductor technology verification competition, Atom was confirmed to be 1.4 to 3.4 times faster in image processing (ResNet) and 1.4 to 2 times faster in language models (BERT-Large) compared to NVIDIA’s A2 and T4 and Qualcomm’s Cloud AI 100. With a maximum 60-watt low-power design and Samsung’s 5-nanometer extreme ultraviolet (EUV) process manufacturing, it achieves about 6 times higher power efficiency for vision models and about 2 times for language models compared to GPUs mainly used in the industry, drastically reducing power consumption.

Rebellions is also preparing an upgraded version of Atom’s performance. In line with this, KT Cloud plans to advance its NPU infrastructure and introduce the next NPU model, ‘Rebel,’ after 2024.

Park Seong-hyun, CEO of Rebellions, said about the NPU infrastructure service, “We take great pride that Atom, which recently demonstrated global top-tier capabilities in both vision and language models at MLPerf, has been commercialized in data centers through KT Cloud’s service. Based on Rebellions’ technology, we will contribute to revitalizing Korea’s AI industry and continue to lead the AI semiconductor market.”

KT Cloud plans to design and build an NPU farm and cloud platform, secure AI semiconductor references through various AI application service demonstrations, and lead innovation in AI infrastructure with low cost, high performance, and high efficiency through diverse technological collaborations. By completing a full AI stack encompassing domestic AI semiconductors, software stacks, cloud platforms, and AI application services, KT Cloud aims to enter the global market by 2025.

Yoon Dong-sik, CEO of KT Cloud, said, “KT Cloud has lowered the barrier to the AI industry by launching the pay-as-you-go AI infrastructure service for super-large AI, HyperScale AI Computing (HAC), and is leading innovation in the AI field with the commercialization of the first cloud-based NPU infrastructure in Korea. We will continue to lead the activation of Korea’s super-large AI industry through AI infrastructure innovation and AI semiconductor advancement.”

This content was produced with the assistance of AI translation services.