Nvidia built a supercomputer with 6,080 graphics accelerators
Let's talk about life
|
January 25, 2022
by
AydarKaribullin
|
1 Reply
|
173 Views
|
1 Reaction
You must be logged in to reply to this topic.
Nvidia built a supercomputer for Meta, formerly Facebook. The supercomputer, designed for AI work, uses a record number of DGX A100 servers.
The supercomputer’s performance for TensorFloat-32 workloads is estimated at 1.895 petaflops. The system, called the AI Research SuperCluster, is already being used to train various AI models.
The supercomputer consists of 760 DGX A100 servers. In total the supercomputer received 6,080 A100 GPUs; a single such accelerator can have 40 or 80 gigabytes of memory onboard.
Although the supercomputer is already operational, it is planned to be fully commissioned closer to the end of 2022. The system will be able to train AI models with more than a trillion parameters.
In the future Meta and Nvidia plan to increase the number of GPUs in the new supercomputer to 16,000.