Dogukan TunaResearchMail

Contents

  • The operation nobody specialized
  • Why any of this is called thermodynamic
  • The math: an Ising lattice, sampled by Gibbs updates
  • The kernel
  • What actually changes for an RL loop
  • Where the kernels actually sit
  • Does it work
  • 1. Is it sampling the right distribution
  • 2. Parity, and what the replicas buy
  • 3. Does the temperature ladder earn its place
  • What this actually fits
  • What comes next

Tags

ThermoRL-CUDA, Ultraresearch, reinforcement learning, thermodynamic computing, Ising model, Gibbs sampling, GPU kernels, CUDA, statistical mechanics, physics simulation

Share

XLinkedIn

Building World's First Thermodynamic RL Engine for Fused Ising-Gibbs Sampling Kernels in CUDA

ThermoRL-CUDA, the engine I build at Ultraresearch: chromatic Gibbs sampling on an Ising lattice and replica-exchange annealing, fused into a CUDA kernel that is itself the RL environment step. From Maxwell's demon and p-bits to the kernel, the oracle that grades it, and the problem classes it fits.

Dogukan Tuna · August 10, 2026 · 28 min read

Building World's First Thermodynamic RL Engine for Fused Ising-Gibbs Sampling Kernels in CUDA
Share:XLinkedIn
← Back