Scherm Brasil, Nvidia, and Positivo Servers & Solutions have partnered for the development of the Jairu supercomputer, to be deployed at the Center for Artificial Intelligence and Machine Learning at the University of São Paulo (CIAAM-USP).

The project aims to strengthen the university's AI research infrastructure by combining domestic hardware with global technologies.

The system comprises Supermicro super servers and 96 Nvidia Blackwell B200 GPUs, making it the largest supercomputer in operation in Latin America with this architecture. Scherm Brasil will be responsible for its integration and deployment, with Nvidia supplying the GPUs, and Positivo Servers & Solutions, Supermicro’s official representative in the country, providing the servers.

The acquisition was carried out through direct import from the United States, managed by Scherm. Its winning bid, valued at 40 million reais ($8 million), offered the best combination of technical capability and lowest price among the five competitors in USP’s bidding process, the university said.

The supercomputer’s assembly was completed within 30 days of the contract signing, solidifying the collaboration among the companies involved in the project.

Positivo Servers & Solutions has already participated in a number of other supercomputing projects in the country. In 2022, it was part of the delivery of the Pégaso supercomputer, one of the most powerful in Latin America, and in 2020, it contributed to the construction of the Dragão system. In previous years, it produced servers at the Manaus factory for the Atlas and Fênix projects, developed in partnership with Atos.

Paulo Fonseca, partner and commercial director at Scherm Brasil, said the Jairu cluster was designed to support advanced artificial intelligence, machine learning, and scientific computing workloads, combining high computing power, parallel storage, and low-latency interconnectivity.

He noted that the system is the result of collaboration between companies with expertise across all layers of supercomputing, from hardware to integration, enabling USP to receive an infrastructure ready for cutting-edge research.

For Marcio Aguiar, director of Nvidia’s Enterprise division in Latin America, the project reinforces academia’s role in training professionals and advancing science. He states that the chip company's participation in the development of Jairu will help to strengthen Brazil’s artificial intelligence research and innovation ecosystem.

Jairu features Supermicro servers utilizing AMD Epyc processors, as well as 12 HGX-standard GPU compute nodes, each equipped with eight Nvidia B200 SXM-5 GPUs, connected via NVLink. Storage is based on a BeeGFS parallel system, with approximately 300TB of usable capacity, supported by a dedicated server and a high-density NVMe JBOD array.

This piece was automatically translated from DCD's Portuguese site and edited by a member of DCD's editorial staff.