Click the blue text above to follow for more exciting content
The story of the chip is turning a new page. Recently, the American startup Normal Computing announced that their CN101 chip has officially been taped out. This is not just an ordinary chip; it is the world’s first “thermodynamic computing chip,” designed specifically for AI and high-performance computing. Its concept is completely different from traditional silicon chips, and is even somewhat “counterintuitive”—transforming noise, which was once avoided, into a helpful computing ally.In traditional computing, noise interferes with signal transmission, leading to errors in results. However, in the CN101, randomness and uncertainty are precisely the weapons. The chip’s components initially exist in a semi-random state, and as the program runs, they gradually reach equilibrium, with the final equilibrium state being the solution to the problem. You might ask, is this computing method reliable? The answer is: particularly reliable for certain applications. Opening a web browser? No. Training AI models? Perfect! Tasks like image generation and model parameter optimization are inherently driven by randomness and probability, making the CN101 perfectly suited for such applications.Even more astonishing is its energy efficiency. Normal claims that when processing matrix operations and probabilistic calculations, the CN101’s energy consumption efficiency can be up to 1000 times better than traditional chips. In other words, training an AI model of the same scale could reduce electricity costs by a factor of ten. This is a “lifesaver” for data centers facing high electricity costs.Normal’s ambitions are also significant. They envision a future where AI servers no longer rely solely on CPUs and GPUs, but instead become a “super computing toolbox”: using traditional chips for deterministic calculations, thermodynamic ASICs for probabilistic calculations, and even mixing in quantum chips to achieve “computational freedom.” Even more impressively, their product roadmap extends to 2028, with plans to venture into high-resolution image processing and complex video generation, continuing to leverage the unique advantages of “noise computing.”Currently, silicon-based chips are approaching physical limits, and the energy consumption pressure on AI data centers is increasing, leading to various “new approaches” emerging. Silicon photonics accelerates data transmission with photons, while quantum computing is known for its immense computational power, but the former is still striving for commercial viability, and the latter is far from large-scale implementation. In contrast, Normal’s thermodynamic chip may enter reality more quickly, becoming a new variable in this computational power race.Who would have thought that the noise once despised by everyone could now be the new ace in the AI world? The emergence of the CN101 may be opening a completely different path for computing, providing a new answer to the energy consumption challenges of artificial intelligence. In the coming years, whether it can disrupt the landscape is something everyone should watch closely.
Chinese chips, my heart, please give a follow to support!
Follow me to gain
more exciting content