A new computing architecture built from conventional transistors could perform certain AI tasks with a fraction of the energy required by today's graphics processing units (GPUs), according to researchers from MIT and Extropic Corp. The proposed system, called a Denoising Thermodynamic Computer Architecture (DTCA), uses controlled randomness to perform probabilistic computations directly in hardware, potentially addressing one of the most pressing challenges facing the AI industry: runaway energy consumption.

Why Is AI Data Center Power Consumption Becoming a Crisis?

The scale of the problem is staggering. According to the International Energy Agency (IEA), global data center electricity consumption is projected to more than double to approximately 945 terawatt-hours by 2030, a figure that slightly exceeds Japan's current annual electricity consumption. In the United States alone, firms are spending more on AI-focused data centers each year than the inflation-adjusted cost of the Apollo program, and by 2030, these facilities could consume around 10% of all electricity produced in the country.

The problem extends beyond electricity. Data centers consumed approximately 66 billion liters of water in 2023, with hyperscale and colocation facilities accounting for 84% of that total. Communities in water-stressed regions like Arizona and Texas are increasingly caught between conservation mandates and data center approvals that drain local resources while concentrating profits among a handful of technology firms.

How Does This New Thermodynamic Computer Architecture Work?

Rather than relying on the deterministic calculations that power conventional processors and GPUs, the DTCA uses probabilistic computing, which performs calculations by manipulating probability distributions. The key innovation borrows concepts from diffusion models, the machine learning techniques behind modern image generators like DALL-E and Midjourney.

Instead of asking one large probabilistic model to represent an entire dataset, the researchers divide the task into a sequence of simpler denoising steps. Each step incrementally transforms random noise into structured data, reducing the computational burden placed on any individual model. The architecture organizes thousands of sampling circuits into arrays implementing sparse Boltzmann machines, which are AI models that learn patterns in data by assigning probabilities to different possible outcomes.

Critically, the proposed system does not depend on exotic hardware components. The researchers designed their system around conventional CMOS transistors, using specially designed transistor circuits to generate programmable random numbers. Those random bits form the foundation of the probabilistic computations performed throughout the chip. To support the feasibility of the hardware, the team fabricated and tested an experimental transistor-based random-number generator, with laboratory measurements showing the circuit behaved as expected and remained robust under simulated manufacturing variations commonly encountered during semiconductor fabrication.

What Do the Performance Benchmarks Show?

The researchers evaluated their architecture using Fashion-MNIST, a relatively simple image dataset frequently employed to evaluate machine learning algorithms. They estimate their architecture could generate images with performance comparable to GPU implementations while requiring approximately 10,000 times less energy per generated sample. This estimate reflects the projected energy consumption of a future hardware implementation rather than measurements from a complete working computer.

The team also explored a hybrid approach combining conventional neural networks with thermodynamic hardware. Using a small neural network to compress CIFAR-10 images into a binary representation before processing them with the probabilistic computer, the researchers found they could achieve performance comparable to a traditional generative adversarial network while using roughly one-tenth as many neural network parameters in the deterministic portion of the system.

How Could This Technology Reshape Computing Infrastructure?

• Specialized Hardware Diversification: For decades, improvements in computing largely depended on scaling general-purpose processors before GPUs became the dominant accelerator for machine learning. Thermodynamic computing represents another attempt to identify workloads that can benefit from specialized hardware grounded in statistical physics rather than conventional digital logic.
• Hybrid System Potential: The hybrid architecture combining conventional neural networks with thermodynamic hardware may ultimately prove more practical than expecting probabilistic hardware to solve every aspect of machine learning independently, suggesting a future where multiple specialized processors work together on different stages of computation.
• Broader Hardware Ecosystem: Alongside quantum computers, neuromorphic processors, photonic accelerators and analog AI chips, thermodynamic computing has emerged as another candidate architecture aimed at improving efficiency for specialized workloads, indicating a shift away from relying solely on GPUs for all AI tasks.

The work addresses what researchers describe as the "hardware lottery," the phenomenon where AI algorithms themselves have been shaped by the available hardware. Different hardware could enable fundamentally different, and potentially more energy-efficient, approaches to machine learning. Rather than attempting to make existing GPU architectures incrementally more efficient, the DTCA proposal suggests that rethinking the underlying computing paradigm entirely may be necessary.

However, significant challenges remain. The study focuses on AI inference rather than the more computationally intensive training phase, and scaling the architecture to larger AI workloads remains unresolved. The researchers acknowledge that the estimates are based on simulations and a tested random-number circuit, not a complete working system. Nevertheless, the work signals that the AI industry is beginning to explore radical alternatives to GPU-centric computing as energy demands threaten to exceed what existing infrastructure can sustainably support.

The timing is critical. As data center power consumption reshapes energy systems and local environments across the United States, with Northern Virginia's data centers now accounting for 28% of Dominion Energy's Virginia electricity sales, the pressure to find fundamentally more efficient computing approaches has never been greater.