On the 14th, Fun Science website reported that American chip startup Cerebras Systems has launched a new 5-nanometer wafer level Engine 3 (WSE-3) chip. The company's official website states that this is currently the fastest running artificial intelligence (AI) chip in the world, doubling the previous record. WSE-3 has 4 trillion transistors, making it the largest computer chip to date, specifically designed for training large-scale AI models, and is also expected to be used in the ongoing construction of the Bald Eagle Galaxy 3 AI supercomputer in the future. The WSE-3 chip consists of 900000 AI optimized computing cores integrated on an 8.5 x 8.5 inch silicon wafer, similar to its predecessor WSE-2. In a press release released on the 13th, the company stated that the power consumption and price of WSE-3 are comparable to WSE-2, but the power is twice as high. WSE-2 includes 2.6 trillion transistors and 850000 AI cores. One of the most powerful chips currently used to train AI models is the Nvidia H200 Graphics Processing Unit (GPU), but the chip only contains 80 billion transistors, which is only 1/57 of the WSE-3 transistor count. The WSE-3 chip will provide power for the under construction Vulture Galaxy 3 supercomputer. This supercomputer will consist of 64 Cerebras CS-3AI systems based on WSE-3 chips, with a potential floating point computing power of 80 billion times per second, making it one of the most powerful AI supercomputers. When the Bald Eagle Galaxy 3 and Bald Eagle Galaxy 1 and Bald Eagle Galaxy 2 systems work together, the entire network's floating-point computing power will reach 160 billion times per second. In contrast, the world's most powerful supercomputer, the "cutting-edge" supercomputer located at the Oak Ridge National Laboratory in the United States, has a computing power of 10 billion times per second. The company claims that the CS-3 system has excellent usability and requires less code to train large AI models compared to GPUs. It will be used to train future AI systems that are 10 times larger than GPT-4 or Google's Gemini. It is reported that the GPT-4 uses approximately 1.76 trillion parameters to train the system, while the CS-3 system can handle AI models with 24 trillion parameters. (Lai Xin She)