Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizen OC_Formula for clue delivery! CTOnews.com Abu Dhabi Technology Holding Group G42 has announced the launch of Condor Galaxy-1 (CG-1), the world's fastest supercomputer with 54 million cores and a processing capacity of 4 exaFLOPS (CTOnews.com Note: 4 trillion times per second), which can be used to train the most advanced artificial intelligence models. The computer, located in Santa Clara, California, is operated by Cerebras, an American artificial intelligence company.

With the development of artificial intelligence technology, the demand for supercomputers is increasing, many enterprises need to use supercomputers to train their own models, and some companies such as Microsoft have built expensive infrastructure and leased it to enterprises for use.

Headquartered in Abu Dhabi, G42 is a holding group in the tech sector that aims to create infrastructure for the future world, partnering with countries and businesses to realize its vision. The company is building the Condor Galaxy family of supercomputer systems to help solve the biggest problems facing humanity today.

Many tech companies have announced huge clusters of thousands of graphics processing units (GPUs), the brains of AI models. But Andrew Feldman, CEO of Cerebras, said,"Distributing a single model across thousands of tiny GPUs requires dozens of people with rare expertise and takes months. "

Instead of using a central supercomputer, Feldman and his team at Cerebras built a set of interconnected AI supercomputers that could dramatically reduce the training time of AI models. According to Cerebras 'approach, building a generative AI model takes only a few minutes and can be done by one person.

Condor Galaxy will have a combined processing power of 36 exaFLOPS, unprecedented in today's computing world.

Each supercomputer is at the heart of the Condor Galaxy family of supercomputers, the first of which was unveiled recently. Cerebras assembled 64 of its flagship CS-2 AI processors to form the CG-1 supercomputer.

AMD's EPYC processor cores power the system, with 54 million AI-optimized compute cores, an overall on-chip memory capacity of 82TB, and bandwidth rates of up to 388Tbps between CS-2 systems.

At 16-bit computing, the system can deliver 4 exaFLOPS of computing power, which is four times faster than the fastest supercomputer in the world today. It can be used to train 600 billion parameter models and can be scaled to support 100 trillion parameter models. By contrast, OpenAI's GPT-4 model was trained using 1.7 trillion parameters.

Cerebras and G42 also plan to introduce two more supercomputers, CG-2 and CG-3, in the U.S. in early 2024, while offering CG-1 as a cloud service to customers.

CG-1 can natively process 50000 tokens without any special software libraries. Programming on the CG-1 eliminates the need for complex distributed programming languages, saving valuable time in distributing workflows on GPUs.

Located at Colovore, a hosting facility in Santa Clara, Cerebras will operate under U.S. law to ensure that the computing power is not exploited by hostile nations.

G42 and Cerebras believe the fastest supercomputers will help tackle challenges such as healthcare, energy and climate change.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.