Scaling AI has always felt like a race against the energy clock. Every advancement in AI models demands exponentially more computing power and with it, exponentially more energy. We recently came across some fascinating developments from Extropic that might just flip this narrative on its head. They claim to have built the world’s first scalable probabilistic computer that can run generative AI workloads using orders of magnitude less energy than traditional GPU-based deep learning.
Why energy is AI’s biggest bottleneck
Extropic predicted a few years back that the biggest barrier to AI’s continued growth wasn’t just algorithmic or data related – it was energy. Right now, almost every new data center worldwide is struggling just to supply the electricity needed to run advanced AI models. Serving complex AI to everyone continuously could consume more energy than humanity can realistically produce.
This sets a sharp boundary on AI’s potential. To push past it, one can either generate more energy at staggering scale, a goal requiring huge infrastructure and national support – or drastically reduce the energy per computation AI consumes. This is where Extropic’s work shines: they’re tackling the puzzle from the hardware and algorithm side, aiming to make AI fundamentally more energy efficient.
Rethinking computing with thermodynamic sampling units
Traditional GPUs excel at deterministic computations, they crunch numbers in rigid, step-by-step ways. But Extropic’s new invention, the Thermodynamic Sampling Unit (TSU), flips this model. Instead of running like a conventional CPU or GPU, these TSUs directly sample from complex probability distributions that underlie generative AI, sidestepping huge matrix multiplications.
How? TSUs harness energy-based models (EBMs), which define probabilities via an energy function. The TSU takes input parameters shaping this function and outputs samples from the distribution it defines. By using a probabilistic computing approach, with highly efficient “pbits” that generate tunable random bits – they radically cut down on the traditionally costly movement of data inside chips.
This local communication-focused architecture means TSUs use much less energy per operation since moving data across chips is a known energy guzzler. Instead of separate memory and compute circuits like GPUs, TSUs combine both seamlessly in a distributed manner minimizing energy spent on communication. It’s a fundamental redesign to match the statistical nature of AI computations, not an adaptation of previous graphics-driven logic.
The energy-efficient future of AI algorithms: the denoising thermodynamic model
Extropic didn’t stop at hardware. They created a new generative AI algorithm, called the Denoising Thermodynamic Model (DTM), inspired by diffusion models but specially designed to run on TSUs. Simulations suggest DTMs on TSUs could be up to 10,000x more energy efficient than current GPU deep learning setups for generative tasks.
Simulations suggest DTMs on TSUs could be up to 10,000x more energy efficient than current GPU deep learning setups for generative tasks.
This is no small feat – it implies thermodynamic machine learning might unlock an entirely new era where AI scales not just with raw power but with incredible power efficiency. And because their Python library thrml lets anyone simulate TSU hardware now, researchers can start exploring and developing algorithms for this new paradigm even before the physical chips become widely available.
What this means for the future of AI scaling
Extropic is aiming to clear one of AI’s biggest roadblocks: energy constraints. If their scalable probabilistic computers live up to their promise, the entire AI landscape could shift. Instead of AI development being shackled by power ceilings and costly data centers, creating and running state-of-the-art AI models may become orders of magnitude cheaper and more sustainable. This doesn’t just open doors for more expansive AI deployment globally, from better drug discovery and improved climate forecasting, to smarter automation and democratized cognitive augmentation – but also invites a rethinking of how computer engineering and AI algorithms co-evolve. The shift from deterministic to probabilistic hardware signals a new chapter where AI is organically baked into the physics of computing itself.
Looking ahead, Extropic’s call for experts in integrated circuit design and probabilistic machine learning to join their push shows how multidisciplinary this revolution will be. And their openness in sharing early prototypes and simulation tools paves the way for a community-driven acceleration of thermodynamic machine learning.
- Energy is shaping AI’s future – we must innovate beyond current hardware to scale effectively.
- Thermodynamic Sampling Units represent a hardware paradigm shift: probabilistic computing instead of deterministic processing.
- The Denoising Thermodynamic Model showcases enormous potential for energy-efficient AI algorithms specifically designed for this new hardware.
- Community engagement and open tools like
thrmlcould spur rapid innovation before commercial chips even ship.
It’s exciting to imagine a future where AI’s raw power isn’t limited by power grids but empowered by completely new ways of thinking about computation. Extropic’s thermodynamic computing approach might just be the key to opening that door. As these ideas and prototypes mature, they could inspire a thermodynamic machine learning revolution that finally scales AI sustainably and profoundly.
