Cerebras Systems, a pioneering company in accelerating generative AI, and G42, a UAE-based technology holding group, have jointly unveiled the Condor Galaxy, the world’s largest supercomputer for AI training. This cutting-edge network consists of nine interconnected supercomputers and introduces a groundbreaking approach to AI compute, significantly reducing AI model training time. The first supercomputer on this network, known as Condor Galaxy 1 (CG-1), boasts an impressive 4 exaFLOPs and is equipped with 54 million cores.
The strategic collaboration between Cerebras and G42 aims to deploy two more supercomputers, CG-2 and CG-3, in the United States by early 2024. This ambitious plan will result in a total capacity of 36 exaFLOPs, revolutionizing the field of AI on a global scale.
The CEO of G42 Cloud, a subsidiary of G42, Talal Alkaissi, expressed excitement about this partnership, combining Cerebras’ exceptional compute capabilities with G42’s multi-industry AI expertise. They envision Condor Galaxy being employed to address society’s most critical challenges in healthcare, energy, climate action, and other domains.
Located in Santa Clara, California, CG-1 is an AI supercomputer formed by linking 64 Cerebras CS-2 systems, making it user-friendly and easy to use. It is offered as a cloud service, enabling customers to access the high-performance AI supercomputer without the complexities of managing or distributing models across physical systems.
CG-1 marks a significant milestone for Cerebras, as it represents their first dedicated AI supercomputer managed and operated in collaboration with G42. The supercomputer is designed to facilitate rapid and straightforward training of large and innovative AI models, thus accelerating advancements in various fields. Already, the Cerebras-G42 partnership has made strides in developing state-of-the-art AI models for Arabic bilingual chat, healthcare, and climate studies.
Andrew Feldman, CEO of Cerebras Systems, highlighted CG-1’s remarkable capabilities, delivering 4 exaFLOPs of AI compute at FP 16, effectively reducing AI training timelines and eliminating the challenges associated with distributed compute. He contrasted it with massive GPU clusters that are expensive to build and challenging to use for single-model distribution. CG-1’s simplicity enables a single person to set up generative AI models in minutes, not months. Additionally, they plan to deploy three more 4 exaFLOP AI supercomputers across the U.S., resulting in a staggering 36 exaFLOPs of efficient, purpose-built AI compute in the next year.
G42, a leading AI and cloud computing company based in the UAE, is spearheading significant digital transformation initiatives worldwide. The UAE’s investments in AI, including the establishment of the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), the world’s first post-graduate university solely focused on AI, demonstrate their commitment to advancing AI technologies.
The partnership between G42 and Cerebras tackles the fundamental challenges of training large models by providing immense compute power, vast datasets, and specialized AI expertise. With the Condor Galaxy supercomputing network, they aim to democratize AI by offering simple access to industry-leading AI compute. Leveraging G42’s work with diverse datasets in healthcare, energy, and climate studies, users of these systems can train cutting-edge foundational models, resulting in powerful applications and positive impacts on various fields. The collaboration’s diverse team of hardware engineers, data engineers, AI scientists, and industry specialists creates a comprehensive AI solution, driving groundbreaking results and catalyzing numerous AI projects worldwide.
