Nvidia Launches GH200 Grace Badminton, A Chip That Can Handle Generative AI Heavy Work

Photo illustration. (doc. Nvidia)

JAKARTA - Nvidia launched the GH200 Artificial Intelligence (AI)-based chip Gracemen, which is based on the new Gracemen Superchip with the world's first HBM3e processor.

Nvidia founder and CEO Jensen Huang said the chip was made for an accelerating era of computing and a growing generative AI.

GH200 Grace fire is designed to handle the most complex generative AI workload, including a large language model (LLM), a recommendation system and a vector database.

The new platform will also be available in a variety of dual configurations, which provide up to 3.5 times more memory capacity as well as its bandwidth, than current generation offerings.

The chip consists of one server with 144 core Arm Neoverse, eight AI performance petaflops, and 282GB of the latest HBM3e memory technology.

"To meet the soaring demand for generative AI, the data center requires an accelerated computing platform with special needs," Huang said in a statement, quoted Wednesday, August 9.

The new GH200 platform Gracemen Superchip'' delivers this with a great memory and bandwidth technology to improve throughput, the ability to connect GPUs to uncompromising aggregate performance, and server designs that can be easily used across data centers," he continued.

Using the Gracemen Superchip, this chip can be linked to an additional Superchip by Nvidia NVLink. They can work together to implement the giant model used for the generative AI.

The high-speed coherent technology gives the GPU full access to CPU memory, providing a combined 1.2TB fast memory while in double configuration.

The memory of HBM3e, which is 50 percent faster than the current HBM3, provides a combined bandwidth of a total of 10TB per second.

"A leading manufacturer already offers a system based on the previously announced Graceten Superchip. To encourage widespread adoption of technology, the next-generation Gracemen Superchip platform with HBM3e is fully compatible with the specifications of the NVIDIA MGX server launched at the COMPUTEX event earlier this year," Huang said.

In MGX, any system manufacturer can quickly and efficiently add Grace fire to more than 100 server variations.

Finally, Nvidia CEO Huang stated that the system running on GH200 Gracemen will begin in the second quarter of (Q2) 2024.