CS-3 AI supercomputer boasts 900,000 cores and 44GB SRAM, rivaling top 10 supercomputers in the world Cerebras Systems' WSE-3 AI chip revolutionizes training with 125 petaFLOPS of power and scalability . Credit: newatlas.com

The release of Cerebras Systems' latest AI chip, the Wafer Scale Engine 3 (WSE-3), has caused quite a stir in the tech world. With its immense power and efficiency, this chip is set to revolutionize the training process of artificial intelligence models.

Some may view AI as a groundbreaking tool with countless benefits, while others see it as a harmful force that only serves the interests of powerful individuals. But regardless of one's perspective, there is no denying that the WSE-3 is a game-changer. Boasting a peak performance of 125 petaFLOPS, it is now the world's fastest AI chip.

What makes the WSE-3 even more impressive is its scalability. This small chip, about the size of a pizza box, contains a staggering four trillion transistors. This is twice the performance of Cerebras' previous model, which was already a world record-holder. And when multiple chips are combined in the Cerebras CS-3 AI supercomputer, the result is a unit that can match the performance of an entire room full of servers in a single mini-fridge-sized system.

According to Cerebras, the CS-3 is equipped with 900,000 AI cores and 44 GB of on-chip SRAM, giving it a peak AI performance of 125 petaFLOPS. This is enough to put it among the top 10 supercomputers in the world. However, since it has not yet been tested on these benchmarks, its actual ranking is uncertain.

In addition to its impressive processing power, the CS-3 also offers a variety of external memory options, ranging from 1.5 TB to a massive 1.2 Petabytes (PB). This allows it to train AI models with up to 24 trillion parameters. To put this into perspective, most AI models currently contain billions of parameters, with the upcoming GPT-4 estimated to have 1.8 trillion.

Furthermore, Cerebras claims that the CS-3 can train a one-trillion-parameter model with the same ease that current GPU-based computers train models with only one billion parameters. This is made possible by the wafer production process used in creating the WSE-3 chips, which allows for scalability. In fact, up to 2,048 units can be clustered together to create a supercomputer capable of a mind-boggling 256 exaFLOPS. For comparison, the top supercomputers in the world currently reach just over one exaFLOP.

This kind of computing power will undoubtedly propel the development of AI models even further. With the CS-3, Cerebras claims that a Llama 70B model can be trained from scratch in just one day. It's no wonder that people are starting to worry about the rapid advancement of AI and its potential impact on various industries and occupations. Whether you are a professional or a hobbyist, it seems that AI systems will be able to outperform human capabilities faster than ever.

Source: newatlas.com
Steven Russell
Steven Russell is a proficient entity from the Technology field. He completed Master’s Degree in Computer Science and Technology. He was engaged in the formation and administration of computational systems in his previous firm. He is associated with Industry News USA from last 2 years. Due to his command over the technology field, he has become the head of the Technology section in very less time period. “Latest gadgets” is the thing that attracts Steven the most.