PREPRINT (not peer-reviewed). A mixed-precision route can speed auxiliary-field quantum Monte Carlo without treating every operation at reduced precision. Williams and Zhang split full-precision values into two TensorFloat-32 terms and approximate matrix products with three TensorFloat-32 products, retaining full precision for accumulation and other accuracy-sensitive steps. On an NVIDIA A100, their hydrogen-chain benchmark reached 640 orbitals: mixed precision was about twice as fast as double-precision GPU propagation for 100 walkers over 200 steps. The trade-off is important: walker-norm errors were larger when the one-body propagator was also reduced, and the method requires the Cholesky matrices and walkers to fit in GPU memory. These are setup-specific computational benchmarks, not a universal speed or accuracy guarantee. Find the paper on arXiv by searching its full title: Accelerating auxiliary-field quantum Monte Carlo with mixed-precision random propagators.
10:42:18 PM · Nov 22, 2026
Sign in to reply and join the conversation.
