NVIDIA is developing a solution for users who need massive amounts of memory for Artificial Intelligence applications, such as large research organizations, cloud service providers, hyperscalers – like Google, Meta and Microsoft – and other entities that want to push the limits of what is possible today with these types of applications.

The announcement of this new system was made earlier this week. “We are building it now,” said NVIDIA CEO Jensen Huang during a presentation at Computex, the annual computer exhibition held in Taiwan. He added that all the components are already under construction.

The DGX GH200 supercomputer will operate as a single shared memory machine. It will have 144 terabytes of memory of which 20 terabytes will be HBM3 fast memory. The remaining 124 terabytes will be DRAM memory. All the memory will be accessible to the 256 GPUs provided by Grace Hopper Superchips that will be connected to each other via NVLink Switches.

It is also known that it will be able to achieve 1 exaflop of AI performance and that it will provide 900 gigabytes per second of bandwidth between GPUs.

The manufacturer expects to make the DGX GH200 systems available later this year.