Cerebras Systems, a leading global company in innovative artificial intelligence technology stimulation field, in partnership with G42, a major stable technological solutions company in Abu Dhabi, has announced the start of the development project of Condor Galaxy 3 (CG3).
This project is the third version in the series of massive AI-supporting computers Condor Galaxy, containing 64 units of the new Cerebras CS-3 system, each equipped with the most powerful AI chip on the market – Wafer-Scale Engine 3 (WSE-3), providing a computational power of 2.6 exaflops, supported by 20 million cores designed specifically to enhance AI performance.
The strategic partnership between Cerebras and G42 has contributed to providing high-level computational power of 8 exaflops by utilizing the supercomputers 1 Condor Galaxy and 2 Condor Galaxy, among the most powerful in the AI field globally. With the addition of the 3 Condor Galaxy supercomputer, based in Dallas, Texas, the total computational power of the Condor Galaxy network rises to 16 exaflops.
Said Kirill Evteimov, CTO of G42 upon the launch of Condor Galaxy:
“We are taking another step forward towards achieving our common goal of revolutionizing the global infrastructure for artificial intelligence by developing the largest and fastest giant computers specifically designed for AI in the world.”
He added:
“The current Condor Galaxy network has played a role in training some of the best open-source models in this field, which have received thousands of downloads. Now, with its capacity boosted to 16 exaflops, we are excited to see the new innovations that we can achieve thanks to the capabilities of these giant computers.”
The third node of Condor Galaxy consists of 64 systems of Cerebras CS-3 systems, each equipped with the third generation chip type 3-WSE, containing 4 trillion transistors and 900 thousand cores dedicated to AI processing.
These chips have been manufactured in TSMC labs using 5-nanometer technology, providing double efficiency in performance compared to the previous generation in terms of power and cost. The 3-WSE is specifically designed for training the largest AI models available in the industry, capable of achieving an astonishing performance rate of up to 125 petaflops per chip.
Meanwhile, Andrew Feldman, CEO and co-founder of Cerebras, expressed his great pride with the following statement:
We take great pride in the pivotal role that our system, named CS-3, will play in the exceptional strategic partnership with G42. The Condor Galaxy 3 system and the upcoming supercomputers will collectively contribute to delivering tens of exaflops of computer capabilities for artificial intelligence, marking a major breakthrough in the world of AI with unprecedented processing power and supreme efficiency.”
Condor Galaxy has succeeded in developing cutting-edge models in the field of generative artificial intelligence in the Jais-138 sector. The models BTLM-3B-8K, Crystal-Coder-7B, Med42, along with Jais-308, are among the world’s leading bilingual Arabic language models. They can now be found on the Azure Cloud platform.
The 8-3-TLM model tops the list of models containing 3 billion parameters on the HuggingFace site, delivering a performance similar to a 7-billion-parameters model but with a much lighter inference-oriented package.
The Med42 model was developed in collaboration with the entities G42 and Core42. It is considered a leading clinical large language model program that was trained by Condor Galaxy over a weekend holiday, surpassing the performance and accuracy of the MedPaLM model. Condor Galaxy 3 is expected to be available in the markets by the second quarter of 2024.
