Insider Brief
- Researchers from Extropic and MIT proposed a thermodynamic computing architecture that uses probabilistic hardware built from conventional CMOS transistors to perform certain AI generation tasks.
- Simulations suggest the architecture could achieve GPU-comparable performance on a simple image-generation benchmark while using roughly 10,000 times less energy per generated sample.
- Although the system remains experimental, the work highlights growing interest in specialized AI hardware beyond GPUs as energy demands from large-scale AI continue to rise.
As artificial intelligence drives an unprecedented expansion of power-hungry data centers, researchers are increasingly looking beyond graphics processing units (GPUs) for new ways to run machine learning models. One of the latest proposals comes from an emerging field known as thermodynamic computing, where controlled randomness—not deterministic calculations—becomes part of the computation itself.
Researchers from Extropic Corp. and the Massachusetts Institute of Technology have proposed a transistor-based computing architecture that they say could dramatically reduce the energy required for certain AI generation tasks. Their study, published in npj Unconventional Computing, describes what they call a Denoising Thermodynamic Computer Architecture (DTCA), a probabilistic computing system that uses conventional CMOS transistors instead of specialized AI accelerators.
Based on simulations and laboratory measurements of a key hardware component, the researchers estimate that a future implementation could achieve performance comparable to GPU-based image-generation models on a simple benchmark while consuming roughly 10,000 times less energy per generated sample.
Although the architecture remains experimental and has not been demonstrated as a complete computer, the work reflects a broader search for new hardware designs capable of supporting AI’s rapidly growing computational demands.
AI’s Growing Energy Challenge
The proposal begins with a problem that has become increasingly difficult for the AI industry to ignore.
“The unprecedented recent investment in large-scale AI systems will soon put a strain on the world’s energy infrastructure,” the researchers write. They note that U.S. companies are investing in AI-focused data centers at a pace exceeding the inflation-adjusted cost of the Apollo program and cite projections suggesting that AI data centers could consume roughly 10% of U.S. electricity generation by 2030.
Most of today’s AI systems run on GPUs, which excel at performing the matrix calculations required by neural networks. But the researchers argue that AI algorithms have largely evolved around the capabilities of GPU hardware—a phenomenon sometimes called the “hardware lottery.” Different hardware, they suggest, could enable entirely different machine learning approaches that prioritize energy efficiency rather than raw computational throughput.
Instead of accelerating existing neural networks, the new architecture rethinks how some AI models perform computation.
Computing With Probability
At the heart of the proposal is probabilistic computing, a computing approach that manipulates probability distributions rather than relying entirely on deterministic calculations.
The architecture is built around Boltzmann machines, a class of AI models that learn patterns by assigning probabilities to different possible outcomes rather than producing a single fixed answer. These models are well suited to generative AI because they can gradually transform random inputs into realistic outputs.
Previous probabilistic computing systems attempted to model entire datasets with a single large energy-based model. As those models became more expressive, however, they also became much harder to sample efficiently, reducing many of the anticipated energy savings.
The new work attempts to avoid that problem by borrowing ideas from diffusion models—the same family of machine learning techniques behind many modern image generators.
Rather than using one large probabilistic model, the researchers divide generation into a sequence of smaller denoising steps. Each stage gradually converts random noise into structured information, allowing each probabilistic model to remain relatively simple while collectively producing increasingly realistic outputs.
According to the authors, this approach overcomes a longstanding limitation that has hindered earlier probabilistic hardware.
An All-Transistor Design
Unlike several previous probabilistic computing proposals that depended on exotic hardware components, the DTCA is designed entirely around conventional CMOS transistor technology.
The researchers developed specialized transistor circuits that generate programmable random numbers. These random bits form the basis of the probabilistic calculations carried out throughout the processor.
Thousands of these sampling circuits are organized into arrays implementing sparse Boltzmann machines. Instead of constructing one large probabilistic model, multiple smaller models work together to progressively refine noisy data into meaningful outputs.
Because the design relies on mature semiconductor technology, the researchers argue that it may ultimately prove easier to manufacture than earlier probabilistic computing concepts that depended on emerging magnetic or photonic devices.
To validate one of the architecture’s key building blocks, the team fabricated and tested an experimental transistor-based random-number generator. Laboratory measurements showed that the circuit produced the expected probabilistic behavior and remained robust under simulated manufacturing variations commonly encountered in semiconductor fabrication.
Benchmark Performance
Because a complete thermodynamic computer does not yet exist, the researchers evaluated their design using GPU simulations informed by measurements from the experimental random-number generator.
Their primary benchmark used Fashion-MNIST, a relatively simple image dataset widely used in machine learning research.
Under those conditions, the simulated architecture produced image quality comparable to GPU-based approaches while requiring an estimated 10,000 times less energy per generated sample. The researchers emphasize that this figure represents projections from a physical energy model rather than measurements from a finished computing system.
The team also evaluated a hybrid design combining conventional neural networks with thermodynamic hardware.
In that configuration, a small neural network compressed CIFAR-10 images into a binary representation before passing them to the probabilistic hardware. The resulting system achieved performance comparable to a traditional generative adversarial network while using roughly one-tenth as many neural network parameters in its deterministic component.
The researchers suggest that hybrid architectures combining conventional AI accelerators with probabilistic processors may prove more practical than expecting thermodynamic hardware to replace GPUs outright.
Toward Heterogeneous AI Infrastructure
The study reflects a broader shift in how researchers envision future AI infrastructure.
For decades, advances in computing largely came from making general-purpose processors faster. More recently, GPUs became the dominant hardware platform for AI training and inference.
Today, however, many researchers expect future AI systems to combine multiple specialized processors, each optimized for particular workloads. GPUs may continue handling dense neural network computations, while other accelerators—including photonic processors, neuromorphic chips, analog AI hardware and thermodynamic computers—could perform tasks better suited to their underlying physics.
Rather than replacing GPUs, these specialized architectures may eventually work alongside them.
Challenges Remain
Despite the promising results, the researchers acknowledge that significant hurdles remain before thermodynamic computing could become commercially viable.
Most notably, the architecture has only been demonstrated through simulations, with the transistor-based random-number generator representing the only experimentally validated hardware component. The benchmark datasets used in the study are also far simpler than today’s frontier AI models.
Scaling remains another open question. Simply increasing the size of the probabilistic models eventually reduces their effectiveness, suggesting additional algorithmic breakthroughs will be required before the architecture can address the complexity of modern large language models or state-of-the-art image generators.
The researchers argue that future progress will likely come from closer integration between probabilistic hardware and conventional neural networks rather than replacing existing AI accelerators altogether.
Even so, they describe the work as an important proof of concept that demonstrates the potential of probabilistic computing to become another specialized processor in tomorrow’s AI infrastructure.
“The broad analysis presented in this manuscript, which spans from high-level algorithmic ideas to laboratory measurements of novel analog circuits, establishes, for the first time, that a probabilistic computing system could substantially outperform traditional AI hardware,” the researchers write. “Taken as a whole, this work presents a compelling case for significant investment in the continued development of low-energy probabilistic computing systems.”
