GPU-First Heisenberg-Picture Tensor Network Dynamics for the 2D Transverse-Field Ising Model
This work accelerates tensor network simulations for condensed matter physicists studying 2D quantum systems.
CppSim achieves a 7.6x Trotter speedup for 2D Ising tensor network dynamics on GPUs via zero-malloc workspaces, custom GPU tensor permutations, and hybrid QR strategies.
We present CppSim, a C++/GPU 2D Ising simulator for Heisenberg-picture tensor network time evolution on GPUs. The key computational contributions are: first, a zero-malloc GPU workspace that pre-allocates all buffers at startup; second, a custom GPU tensor permutation kernel replacing host-side index shuffling with a pure device-to-device operation, yielding a 7.6x trotter speedup; third, a hybrid QR strategy selecting Cholesky-QR for tall-skinny matrices and Householder-QR otherwise; fourth, adaptive Belief Propagation with log-space Bethe partition function evaluation and explicit sign tracking.