The convergence line
A 539-atom Cantor alloy in NVIDIA’s PET experiment exposes a distinction that runtime can easily hide. The FIRE2 optimizer changes atomic positions until the forces become sufficiently small. The FP32 baseline reaches the target in 218 steps; TF32x3 repetitions finish in 218 to 222 steps. Single-pass TF32 does not finish within the 500-step limit. Raising the limit to 1000 still does not produce convergence. A shorter step on a path that never reaches the researcher’s endpoint does not deliver a completed calculation.[1]
My preferred unit of comparison here is the complete calculation reaching the same force tolerance. The company tests a threshold of 0.001 electronvolt per angstrom. Conditioning runtime on that endpoint asks how well the numerical method preserves the scientific task. Work per step and a completed structural relaxation are different measurements. Their separation in PET makes the choice of endpoint consequential. For the researcher, a usable outcome includes that stopping condition as well as the time spent computing.[1]
The path through precision
PET is a model that learns interatomic interactions. The new technical study constructs matrix products from lower-precision components and combines them into an FP32 result. TF32x3 does not follow the same computational path as single-pass TF32. Inputs, weights and outputs remain FP32, and retraining is unnecessary. This design allows a change in numerical execution to be tested without changing the model’s learned weights. The comparison concerns numerical execution of the same model, rather than the discovery of a new material.[1]
My inference is that failure in the alloy experiment is consistent with small force inconsistencies accumulating across iterations. The optimizer repeatedly uses forces predicted by the model, and convergence is measured at the end of that chain. A plausible alternative is that this alloy, starting structure and FIRE2 configuration are particularly unfavorable to the single-pass method. Repeating the separation with other potentials and starting structures would provide stronger support for generalization.[1]
Adding physical endpoints
Another test follows a 192-atom water system for 30 picoseconds. In this NVE experiment, where total energy is expected to remain conserved, FP32 and TF32x3 show bounded energy fluctuations while single-pass TF32 exhibits systematic drift. A separate calculation compares energy-curve smoothness as carbon monoxide moves away from a platinum surface. The tests therefore examine more than a difference in one force prediction: they include a trajectory that repeatedly uses those forces and an energy curve across positions.[1]
Agreement across these endpoints increases numerical confidence in TF32x3 under the PET conditions. The evidence comes from NVIDIA’s technical experiments; it does not include an independent replication across all interatomic models. A stronger next test is for another team to compare complete calculation time with a different potential while retaining the same force tolerance and energy criteria. I prefer to assess acceleration within that design: runtime acquires meaning together with a completed physical calculation.[1]