IntroducingTHRML
THRML is a library to construct thermodynamic hypergraphical models, a fabric of probabilitic models made to be run efficiently on TSUs.
Algorithm Development
Simulate TSUs
thrml uses JAX to massively accelerate simulations of PGMs, including the Energy-Based Models that will be instantiated on our Z1 chip.
We use thrml to run simulations of our hardware running our algorithms.
Hardware Design
Explore Chip Architectures
thrml can be used to experiment with new TSU architectures, changing the connection graph, node type, and other design parameters.
We use thrml to identify promising future hardware designs.
Machine Learning
Train Energy-Based Models
thrml can compute gradients in probabilistic graphical models, which can be used to train energy-based models.
We used this approach to develop Denoising Thermodynamic Models.
OPEN RESEARCH
We're offering research grants to support innovative work in probabilistic computing. We welcome applications from all qualified researchers, particularly early-career researchers and PhD students.
PARTNERSHIPS
We are collaboratively developing applications of thermodynamic computing systems by partnering with companies that run large probabilistic workloads. If your organization is interested in pioneering the future of computing, get in touch.
Diffusion Models
Text, image & video
Simulations
Nuclear, chemical & molecular
World models
Weather, robotics, autonomy
Optimization
Discrete or continuous, grey box
Probabilistic Inference
Bayesian filtering, financial modeling
