In a groundbreaking development in the world of artificial intelligence, the World’s Fastest AI Chip has been unveiled by Cerebras, boasting an impressive 4 trillion transistors. The new WSE-3 chip delivers an outstanding performance, surpassing its predecessor, the Cerebras WSE-2, by twice the capacity while maintaining the same power consumption and cost efficiency.
Designed specifically for training the largest AI models in the industry, the WSE-3 chip, based on 5nm technology, powers the Cerebras CS-3 AI supercomputer. With a remarkable 125 petaflops of peak AI performance, this supercomputer features 900,000 AI optimized compute cores, ensuring unparalleled processing power.
One of the most striking features of the CS-3 is its massive memory system, capable of storing up to 1.2 petabytes. This cutting-edge technology enables the training of next-generation models that are ten times larger than GPT-4 and Gemini. The CS-3 can accommodate a model with 24 trillion parameters in a single logical memory space without the need for partitioning or refactoring, streamlining the training process and significantly boosting developer productivity.
The versatility of the CS-3 extends to both enterprise and hyperscale requirements. From fine-tuning 70B models in a day with just four system configurations to training Llama 70B from scratch in a single day using 2048 systems, the CS-3 caters to a wide range of AI applications with unparalleled speed and efficiency.
This revolutionary AI supercomputer represents a significant leap forward in the field of artificial intelligence, offering unmatched performance and scalability for training the most complex AI models. The launch of the Cerebras CS-3 is poised to accelerate advancements in AI research and development, opening up new possibilities for innovative applications across various industries.
