Cerebras Systems and technology holding group G42 have announced the Condor Galaxy project. This is his network of 9 interlinked supercomputers with a total performance of 36 FP16 ExaFLOPs for training AI models. The first supercomputer named Condor Galaxy 1 (CG-1) boasts 4 ExaFLOPs of his FP16 performance and his 54 million cores. CG-2 and CG-3 will be installed in the United States and will continue in 2024. The remaining systems will be installed around the world, bringing the total cost of the project to more than $900 million.
The CG-1 supercomputer, located in Santa Clara, California, integrates 64 Cerebras CS-2 systems into a single user-friendly AI supercomputer, providing 4 ExaFLOPs of dense, systolic Can provide FP16 computing. The machine is based on Cerebras’ 2.6 trillion-transistor second-generation wafer-scale engine processor and is specifically designed for large-scale language models and generative AI. It supports up to 600 billion parameter models and has configurations that scale to support up to 100 trillion parameter models. Cerebras says 54 million AI-optimized computing cores and 388 Tb/s of massive fabric network bandwidth deliver nearly linear performance from 1 to 64 of his CS-2 systems can be extended.
The CG-1 supercomputer also offers inherent support for long sequence length training (up to 50,000 tokens) without the need for complex distributed programming languages typical of GPU clusters.
“Achieving 4 exaflops of AI compute in FP16, CG-1 dramatically shortens AI training timelines while eliminating the pain of distributed computing,” said Andrew Feldman, CEO of Cerebras Systems. I’m here. “Many cloud companies have announced large GPU clusters that cost billions of dollars to build, but are very difficult to use. Distributing a single model to thousands of small GPUs requires Dozens of staff with rare expertise take months, CG-1 eliminates this challenge Generative AI model setup takes minutes instead of months and can be run by one person CG-1 is the first of three 4 ExaFLOP AI supercomputers to be deployed across the U.S. Over the next year, we will work with G42 to scale this deployment to a staggering 36 exaflops. We plan to launch a highly efficient purpose-built AI computing platform.”
The supercomputer is provided as a cloud service by Cerebras and G42, and Cerebras and G42 claim that because it is located in the United States, it cannot be used by hostile nations.
CG-1 is the first of four FP16 ExaFLOP AI supercomputers (CG-1, CG-2, and CG-3) co-created by Cerebras and G42 and located in the United States. When these three AI supercomputers are connected, they form 12 AI supercomputers. FP16 ExaFLOP, a distributed AI supercomputer with 162 million cores, but it remains to be seen how efficient this network will be.
In 2024, G42 and Cerebras will launch six more Condor Galaxy supercomputers worldwide, increasing the total computing power to 36 FP16 ExaFLOPs delivered by 576 CS-2 systems. increase.
The Condor Galaxy project aims to democratize AI by providing advanced AI computing technology in the cloud.