Researchers and academic institutions can now run large AI models, simulations, and data-intensive experiments efficiently with a high-performance supercomputer.
Artificial intelligence research is increasingly limited due to limited computing power. Training large models, running complex simulations, and processing massive datasets require high-performance systems. In Latin America, academic institutions have faced limitations in infrastructure, slowing AI innovation and constraining the development of models tailored to local contexts. Researchers needed scalable environments capable of supporting data-intensive AI, machine learning, and scientific computing workloads.
The Jairu supercomputer addresses this need. Equipped with 96 NVIDIA B200 GPUs and Supermicro servers, it allows researchers to run AI models, perform simulations, and process complex algorithms efficiently. The system integrates high-capacity storage and low-latency interconnects, reducing bottlenecks and enabling large-scale experiments.
Its architecture provides several advantages. Head nodes manage the environment, including login, orchestration, and workload distribution. GPU nodes follow the HGX standard and use NVLink for fast inter-GPU communication. Storage and networking are designed to maintain high throughput and low latency. The system also includes software for centralized management, GPU scheduling, and experiment reproducibility.
Key features of the supercomputer include:
- 5 head nodes handle management, login, and orchestration.
- 12 GPU nodes each have 8 NVIDIA B200 SXM-5 GPUs.
- NVLink connects GPUs for fast communication.
- Storage uses BeeGFS with around 300 TB usable.
- Networking combines Ethernet 200 Gb/s and InfiniBand 800 Gb/s.
- NVIDIA AI software manages workloads and schedules tasks.
Fabio G. Cozman, USP’s Center for Artificial Intelligence and Machine Learning (CIAAM-USP), says, “With Jairu, we have an AI infrastructure that will allow us to develop large models and deepen research relevant to the Brazilian context.” By combining high-performance hardware, enterprise software, and specialized engineering, the system strengthens Brazil’s academic capacity and competitiveness in AI and high-performance computing.
