- Normal Computing announces CN101, the world’s first thermodynamic computing chip
- The startup says its approach supports scaling AI workloads within current data center power limits
- Future designs are intended to deliver higher performance inside existing infrastructure
Normal Computing has announced the successful tape-out of CN101, which it describes as the “world’s first thermodynamic computing chip.”
The startup sees this development as a natural response to the mounting energy demands of AI and scientific workloads.
Unlike CPUs and GPUs that rely on deterministic logic, CN101 is designed to exploit natural dynamics such as fluctuations, dissipation, and randomness.
Normal’s promising roadmap
The idea is to accelerate certain reasoning tasks while lowering energy use by drawing on processes that existing chips typically suppress.
The company says CN101 is targeted at two specific categories of computation. One involves large-scale linear algebra, which is central to optimization problems and scientific modeling.
The other is stochastic sampling, where Normal’s lattice random walk approach is intended to speed up statistical methods, including Bayesian inference.
“In recent months, we have seen that AI capabilities are approaching a flattening curve with today’s energy budgets and architecture, even as we plan to scale training runs another 10,000x in the next 5 years. Thermodynamic computing has the potential to define the next decades’ scaling laws by exploiting the physical realization of AI algorithms, including post-autoregressive architectures. Achieving first silicon success is a historic moment for this emerging paradigm – executed by a radically small engineering team,” said Faris Sbahi, CEO at Normal Computing.
Looking ahead, the company has set a roadmap that begins with CN101 but stretches into the next decade.
“Our vision to scale diffusion models with our stochastic hardware starts with demonstrating key applications on CN101 this year, then achieving state-of-the-art performance on medium-scale GenAI tasks next year with CN201, and finally achieving multiple orders-of-magnitude performance improvements for large-scale GenAI with CN301 two years from now.” Patrick Coles, Chief Scientist at Normal Computing explained.
Normal engineers say the tape-out also represents the first step toward characterizing how these ideas behave in real silicon.
“CN101 represents the first silicon demonstration of our thermodynamic architecture that leverages randomness, metastability, and noise to perform sampling tasks. By characterizing CN101, we’ll be able to lay the groundwork for understanding how these random processes behave on real silicon, and chart a clear path towards scaling up our architecture to support state-of-the-art diffusion models,” Zach Belateche, Silicon Engineering Lead at Normal said.
Normal Computing was founded in 2022 by engineers from across Google Brain, Google X, and Palantir.