NVIDIA and Foxconn will produce Taiwan’s fastest supercomputer

NVIDIA and Foxconn will join forces to produce Taiwan’s fastest artificial intelligence supercomputer, powered by Blackwell GB200 NVL72 servers.

The platform is the most powerful of the entire Blackwell architecture, being used to further boost the country’s technology industry through the collaboration of the two manufacturers – using Grace CPUs and new generation server GPUs.

According to Foxconn and NVIDIA’s announcement, it will deliver 90 exaflops of artificial intelligence performance – the fastest Taiwan will have across its entire territory. This will also help the country to position itself in the industry as one of the main leaders in this market.

The objective is to center in the country some advanced operations such as scientific research such as that carried out to study cancer, development of language models and even “smart city” innovations for metropolitan areas.

See also:

Union between NVIDIA and Foxconn

Foxconn vice-president, James Wu, says that the supercomputer they are producing alongside NVIDIA will be one of the fastest on the planet and will bring several innovations to Taiwan.

Powered by the NVIDIA Blackwell platform, Foxconn’s new artificial intelligence supercomputer is one of the most powerful in the world, representing a leap forward in AI computing and efficiency

James Wu

Foxconn’s plan is to bring focus to three areas: manufacturing, cities and electric vehicles. The proposal is to use the supercomputer to allow artificial intelligence to assist services in urban areas.

The power of NVIDIA hardware

In terms of hardware, each NVIDIA GB200 NVL72 rack will have 36 Grace processors and 72 GB200 Blackwell GPUs – connected via an NVLink domain.

In other words, each rack of the NVIDIA and Foxconn supercomputer will have 3,240 TFLOPS of FP64 and Tensor Core FP64 performance, capable of reaching up to 13.5TB of HBM3E memory and a memory bandwidth of 576 TB/s.

Compared to NVIDIA H100 GPUs, the GB200 NVL72 can bring 30x more LLM inferences, 4x better LLM training, and up to 18x better data processing through the CPU – which is 25x more powerful and efficient than the H100 .

Taking into account that the NVIDIA and Foxconn supercomputer built in Taiwan will use the GB200 NVL72 – which has 64 racks – it will have a total of 4,608 high-performance GB200 Blackwell GPUs.

Fonte: WCCFTECH