NVIDIA announced that it will join forces with partners toThe U.S. government is building seven AI supercomputers, all of which can reach EFlops scale, serving Aragon National Laboratory and Los Alamos National Laboratory under the U.S. Department of Energy.
Among them, the first two units are composed ofOracleConstruction, for Aragon National Laboratory.
The first one is called "Equinox" and is equipped with 10,000 Blackwell GPUs. The first phase of the project will go online in 2026.
The second one is called "Solstice", which can also be said to be the second phase of the project.The number of Blackwell GPUs used will exceed 100,000, and the power consumption will reach 200 megawatts, which is 200 million watts.
Taken together, the total computing power of FP4 is as high as 2200EFlops, which is 22 trillion floating points per second!
It can perform both classical scientific calculations andSupports AI large model simulation with up to 3 trillion parameters.
It is worth mentioning that the choice of Oracle for this cooperation is quite unexpected, because today's Oracle's core business is cloud infrastructure that supports AI and high-performance computing, unlike Atos, Dell, Hewlett Packard Enterprise, etc., which can customize designs and build supercomputers for customers.
Therefore, it remains to be seen how Oracle will build such a large-scale supercomputer, and it is not clear whether Oracle will manage it after it goes online.
The other five supercomputers have only announced their names. “Tara”, “Minerva” and “Janus” are all based on Blackwell and are also used in the Aragon National Laboratory.
"Mission" and "Vision" will switch to the next generation Rubin, for Los Alamos National Laboratory.
