TSU 101 an New Type of Computing Hardware
7 days ago
- #Probabilistic Computing
- #AI Hardware
- #Energy Efficiency
- Extropic is developing Thermodynamic Sampling Units (TSUs), a new type of energy-efficient computing hardware for generative AI.
- TSUs are probabilistic computers designed to sample from probability distributions, making them inherently suited for generative AI tasks.
- Unlike GPUs, which excel at matrix multiplication, TSUs are optimized for sampling, a core operation in generative AI algorithms.
- The hardware leverages probabilistic circuits called pbits, which can be combined to sample from complex distributions using Gibbs sampling.
- Extropic's approach avoids the inefficiencies of traditional energy-based models by using Denoising Thermodynamic Models (DTMs), inspired by diffusion models.
- The upcoming Z1 TSU chip will feature hundreds of thousands of sampling cells, enabling scalable and energy-efficient generative AI applications.
- Extropic provides a Python library, thrml, for simulating TSU architectures and developing new thermodynamic algorithms.