Nvidia, SK Telecom and Rebellions signed a three-party memorandum on 9 November to develop a mixed-chip AI computing system for SK Telecom’s data centres, Yonhap reported. The agreement was signed at AI Day Seoul, Nvidia’s two-day technology event in southern Seoul. In a release dated 8 November, Nvidia said the proposed system would connect Rebellions’ Korean-designed processors with its rack infrastructure for AI inference.
Key points
- The companies signed a memorandum to develop a system combining Nvidia GPUs and Rebellions NPUs for SK Telecom’s AI data centres.
- Rebellions plans to connect its next-generation R200 processors through Nvidia’s NVLink Fusion programme and develop systems using its MGX rack architecture.
- The planned inference work assigns response preparation to Nvidia GPUs and response generation to Rebellions NPUs.
- Korea’s Ministry of Science and ICT plans to support similar cooperation with other domestic NPU developers.
Rebellions plans R200 connection through NVLink Fusion
NVLink Fusion is Nvidia’s programme for bringing processors designed by other companies into its rack-scale infrastructure. Under the memorandum, Rebellions plans to connect its next-generation Rebel R200 neural processing units, or NPUs, to that infrastructure and use Nvidia’s MGX rack architecture to develop the system.
Rebellions is developing the R200 for inference workloads. Nvidia says the processor is being designed for low-latency, energy-efficient work with large AI models, including multitrillion-parameter mixture-of-experts models. NVLink networking is intended to connect Rebellions’ processors within a rack, while the MGX designs give the company access to suppliers of rack components and systems.
The proposed system would put locally designed NPUs alongside Nvidia GPUs in SK Telecom’s AI data centres. Korea’s Ministry of Science and ICT said the arrangement could give domestic chip developers a way to test their technology for a wider market, Yonhap reported. The companies’ signed document is a memorandum to cooperate on the system; SK Telecom’s deployment is described as a plan.
SK Telecom plans mixed-chip inference
SK Telecom plans to deploy the systems in its AI data centre and connect them to its own foundation models and public-facing services, according to Nvidia. Those models include A.X K2, while the services include work under Korea’s “AI for All” project. SK Telecom has previously worked with Nvidia on AI development and infrastructure.
The companies plan to split inference between the two kinds of processor. Nvidia GPUs would handle prefill, the stage that processes a prompt before an answer is produced, and Rebellions NPUs would handle decode, which generates the answer. Nvidia says the companies plan to develop that arrangement to improve throughput and power efficiency.
MGX provides shared rack-design elements for GPU and NPU systems, including their footprint, power delivery and cooling, Nvidia says. SK Telecom plans to use that architecture in its data centre. Rebellions, meanwhile, is responsible for developing the R200 and its associated system. Each step is part of the companies’ proposed deployment rather than a reported operating installation.
Rebellions chief executive Sunghyun Park said the companies intend to deliver low-latency, energy-efficient inference by combining its chips and system design with Nvidia’s interconnect technology and SK Telecom’s models and data-centre operations. SK Telecom chief executive Jung Jai-hun said the operator plans to link the resulting infrastructure to A.X K2 and its public services.
Korean ministry seeks more NPU partners
Korea’s Ministry of Science and ICT supports the collaboration. Its stated aim is to connect domestic semiconductor development with the infrastructure used to run AI services. The ministry has also been working to develop the country’s AI chip sector.
Other chip partnerships are being pursued in Korea. AI Affairs reported AMD chief executive Lisa Su’s plans for talks with South Korean AI chipmakers on 8 October. The Nvidia arrangement gives Rebellions a proposed route into rack systems operated by SK Telecom, subject to the companies developing and deploying them.
The ministry plans to support extending cooperation with Nvidia beyond Rebellions to other domestic NPU companies, Chosun reported.
