NVIDIA and Foxconn are collaborating to construct Taiwan’s most formidable AI supercomputer, leveraging the advanced capabilities of the Blackwell GB200 NVL72 servers. This cutting-edge supercomputer will be based in Kaoshiung and will mark Taiwan as a leading player in AI technology.
Foxconn, Taiwan’s renowned electronics manufacturer, has officially announced this ambitious project in partnership with NVIDIA. This latest venture stems from NVIDIA’s groundbreaking Blackwell architecture, designed to propel the supercomputer’s incredible performance capabilities.
According to recent announcements, the planned supercomputer will deliver over 90 exaflops in AI performance, a groundbreaking achievement for Taiwan. This leap in technological advancement will elevate the nation to the forefront of AI-driven innovation on a global scale. The new system is expected to revolutionize sectors such as cancer research, the development of large language models, and smart city infrastructure.
James Wu, Vice President and Spokesperson for Foxconn, expressed the significance of this development, highlighting it as a monumental step forward in AI computing and efficiency. The integration of NVIDIA’s Blackwell platform means this supercomputer will rank among the world’s most powerful systems.
Foxconn’s strategy includes a three-platform approach focusing on manufacturing, urban development, and electric vehicles. The new supercomputer will play a crucial role in advancing AI-assisted urban services, with the potential to transform cities into smarter, more efficient environments. Utilizing the GB200 NVL72 platform’s capabilities, it will operate with up to 36 Grace CPUs and 72 Blackwell GB200 GPUs per rack, interconnected in a robust 72-GPU NVLink domain.
Per NVIDIA’s specifications, each rack is designed to deliver up to an impressive 3240 TFLOPS of FP64 and FP64 Tensor Core performance, along with 13.5 TB of HBM3e memory, providing a remarkable 576 TB/s of memory bandwidth. Compared to its predecessor, the NVIDIA H100 Tensor Core GPU, the GB200 NVL72 offers substantial improvements: 30x higher LLM inferences, 4x more efficient LLM training, and 18x enhanced data processing through the CPU, while being 25x more powerful overall.
The integration involves 64 racks in total, culminating in 4608 Blackwell GB200 GPUs, making it exceptionally suitable for training extensive AI models and executing complex inference operations involving trillion-parameter datasets.
Foxconn aims to introduce this supercomputer by mid-2025, with full deployment anticipated by 2026. The system will integrate several NVIDIA technologies, including the Omniverse, Isaac robotics platforms, and Digital Twins. These technologies will enrich the manufacturing process by offering advanced tools and libraries tailored to enhance efficiency and capability.
This partnership between Foxconn and NVIDIA signifies a groundbreaking advancement in Taiwan’s tech landscape, striking a powerful chord in the future of AI technology.
