Chipmaker Cerebras Systems has completed Condor Galaxy 1, its first of nine supercomputers for AI training. Now up and running, Condor Galaxy 1 cost more than $100M to build. The supercomputer, located in Santa Clara, California, currently boasts a capability of 2 billion operations per second (2 exaflops) and will expand to 4 exaflops soon. - For perspective, an exaflop computer's capabilities equal performing one calculation per second for over 31 billion years.
- It will double in size from 32 Cerebras CS-2 computers to 64 within 12 weeks.
- Cerebras plans to build two more supercomputers in Asheville, North Carolina (CG-2), and another in Austin, Texas (CG-3), with completion targeted for H1 2024.
- It will continue to add more Condor Galaxy installations until it forms a network of nine supercomputers, reaching a total capability of 36 exaflops.
G42 contract: - UAE-based AI company G42 owns Condor Galaxy, using the system to train an Arabic version of OpenAI's ChatGPT.
- The holding company oversees nine AI-based businesses, including cloud provider G42 Cloud.
- Cerebras operates the supercomputers and can rent out any excess capacity commercially as a service.
|
