en.Wedoany.com Reported - In response to the massive energy demands of AI-focused data centers, researchers from Extropic Corp. and the Massachusetts Institute of Technology (including quantum information scientist Isaac Chuang) have proposed a Denoising Thermodynamic Computer Architecture (DTCA), claiming it can perform specific AI tasks at a fraction of the energy consumed by traditional hardware. The findings were published in npj Unconventional Computing.
The study aims to address a core challenge in the AI industry. The team noted that investments in large-scale AI systems are placing immense pressure on global energy infrastructure, projecting that by 2030, AI-focused data centers could consume approximately 10% of total U.S. energy. With limited efficiency gains in traditional GPU (Graphics Processing Unit) architectures and AI algorithms potentially constrained by existing hardware, finding alternative computing pathways has become critical.
The proposal designs a probabilistic computer architecture based on conventional CMOS transistors, leveraging controlled randomness to perform probabilistic computations directly in hardware. Its principle draws from diffusion model concepts, decomposing complex probabilistic modeling tasks into a series of simple denoising steps that gradually transform random noise into structured data, circumventing previous limitations of probabilistic hardware in the "mixing-expressivity trade-off." The hardware core consists of specially designed transistor circuits for generating programmable random numbers, with these random bits forming the basis for on-chip probabilistic computation. A modular array implements multiple cascaded sparse Boltzmann machines. This modular design can be realized through multiple dedicated hardware blocks on a single chip or by multiple chips communicating to execute different stages of computation. The team has fabricated and tested experimental transistor-based random number generators, which demonstrated robustness under simulated manufacturing process variations.
To validate performance, researchers simulated the architecture using GPUs and incorporated measured data from physical random number generators. Benchmark tests on the Fashion-MNIST image dataset showed that the architecture produced image quality comparable to GPU implementations, but with estimated energy consumption only one-ten-thousandth of that per generated sample. Additionally, a hybrid approach combining traditional neural networks with thermodynamic hardware showed promise on the CIFAR-10 dataset, using only one-tenth the number of neural network parameters compared to traditional generative adversarial networks, potentially enabling more practical distribution of computational tasks across different subsystems.
Despite encouraging results, the research team acknowledged limitations. Currently, only the transistor-based random number generator has been physically validated; the complete computing architecture remains in the theoretical simulation stage. The image datasets used for benchmarking are far less complex than modern large language models or advanced generative models. Efficiently scaling the system to handle more complex data remains a core challenge for probabilistic computing. The team believes future progress is more likely to rely on deep integration of probabilistic hardware with traditional neural networks, rather than completely replacing existing AI accelerators. This research should be viewed as a "first step" warranting further investment and exploration.
