Normal Computing announced that the world's first hot-science computing chip "CN101" has successfully completed the Tape-out. This ASIC, designed for AI and high-performance computing (HPC) data centers, is different from traditional s...
Normal Computing announced that the world's first hot-science computing chip "CN101" has successfully completed the Tape-out. This ASIC, designed for AI and high-performance computing (HPC) data centers, is different from traditional silicon-based computing methods. It uses heat learning (and other physical principles) to achieve unmatched computing efficiency on traditional chips.
Normal Computing points out that the CN101 chip focuses on efficient solution to linear generation and matrix calculations, and uses Normal's unique sampling system to handle other probability calculations. The architecture is designed to accelerate computing tasks and can achieve an energy efficiency of 1,000 times in computing load by leveraging the internal dynamics of the physical system.
The heat-based chip is completely different from traditional computing methods, and is closer to quantum computing and chance computing. Message is a big enemy in traditional electronics, but in thermal and probability wafers, it can be used to solve problems.
Zachary Belteche, head of silicon engineering at Normal Computing, recently interviewed by IEEE Spectrum, "We focus on algorithms that can utilize messy, randomness and indeterminateness. The application space of this type of algorithm is very broad, covering fields from scientific computing, AI to linear generation, and so on."
Foreign media "IEEE Spectrum" explains that the thermal chip component will first be in a semi-random state, and then enter the program. When the equilibrium is reached between the components, the system will read the equilibrium state as the calculation result. This calculation method only applies to applications involving indeterminate results, so it is not used to open web browsers, but it can achieve great advantages for tasks such as AI image generation and other training efforts.
Therefore, rather than the CPU and GPU consume a lot of energy to maintain deterministic logic, Normal's chips use randomness to accelerate AI reasoning. Compared with the traditional method, this is extremely powerful in computing efficiency.
Normal Computing's long-term vision is to construct a heterostructure computing server that integrates CPUs, GPUs, heat learning ASICs, probability chips, and even quantum chips to help integrate various components that are most suitable for different problems in AI training servers.
Normal Computing's CN series blueprint includes subsequent versions for 2026 and 2028, expanding to deeper and more commonly used photo and video expansion model applications.
World’s first ‘thermodynamic computing chip’s reach tape out — Normal Computing’s physics-based ASIC changes lanes to train more AI Normal Computing Announces Tape-Out of World’s First Thermodynamic Computing Chip Extended reading: Advent opens USD 1.3 billion to purchase Swiss chip manufacturer U-blox New model R2 is the main reason for delaying the delay! DeepSeek tried to fail because the chip was still up to NVIDIA
