Skip to main content
Back to timeline
arXivSource publication:

Quantum chemistry stops bending tasks to general software: Born builds a program per task and finishes a 50,024-basis-function Hartree–Fock run in 1.41 hours on one node

Synopsis

General-purpose quantum chemistry packages trade single-task efficiency for coverage across molecules; Born instead predicts each strategy's error and runtime before any target-scale run, then has coding agents synthesize a task-specific program. For a Hartree–Fock calculation with 50,024 basis functions, Born finishes in 1.41 hours on a single CPU node, against an estimated 109 hours for a tuned general-purpose baseline, and remains 9.8 times faster once 9.7 hours of development are included.

Interpretation

A quantum chemistry calculation no longer has to run through a general program written for every task; a program can be generated for the single task at hand, bringing large systems that exceeded practical resource budgets back within reach. General-purpose packages fix algorithms, parameters and schedules into a few preset schemes and cannot exploit the molecule, basis set, accuracy requirement and hardware that each task already fixes; for large molecular systems their time or memory demand can exceed practical budgets. On a Hartree–Fock task with 50,024 basis functions, Born finishes in 1.41 hours on a single CPU node, some general-purpose packages fail to complete the task, and a tuned baseline able to run it is estimated to need 109 hours.

The approach decides before execution which strategy both meets the error tolerance and runs fastest, then compiles the selected strategy into a program fitted to that task. Coding agents make programs cheap to write but not efficient ones easy to find: the strategy space is vast, feedback requires costly target-scale runs, and small-scale timings hide target-scale bottlenecks. Fitted up to C51H104, the error model predicts the held-out C86H174 to within 2% for all five approximations; the runtime model stays within 8% of measurement at every held-out size, while a proxy-fitted power law on the same timings is off by 36%, 10% and 58%.

Strategy selection alone yields about a 19-fold speedup, and agent-driven compilation cuts total time from a predicted 5.6 hours to 1.41 hours measured end to end, with error unchanged. The uncompiled program implements the selected strategy with general-purpose kernels and solvers; compilation changes only how the calculation is computed, not what is computed, so the numerical error stays fixed. The compiled run reports a total-energy error of 0.053 millihartree, below the 0.1 millihartree tolerance; development takes 9.7 hours, and adding one run still leaves roughly a 9.8-fold speedup over the tuned baseline.

The same prediction-and-compilation flow cuts the GPU requirement for a biomolecular calculation from 64 to one and shortens the B-DNA duplex from an estimated two months to 8.1 hours. The ubiquitin calculation previously needed Mako spread over 64 GPUs to reach about 1 hour, and the general-purpose route for the B-DNA duplex was predicted at about 1,400 hours. Ubiquitin completes in 2.1 hours on a single node with one A100 and the B-DNA duplex in 8.1 hours, with measured errors of 5.3 and 3.0 millihartree, both below tolerance.

Perspective

The demonstrated scope is single-node Gaussian-basis Hartree–Fock self-consistent-field calculations on a CPU platform and a CPU–GPU platform with one A100, at an error tolerance of 0.1 millihartree; the alkane, ubiquitin and B-DNA duplex tasks stand for chain, globular and double-helix systems respectively.

The predicted error is extrapolated and compilation leaves it unchanged, and the final program runs once at the target scale to measure convergence, runtime and peak memory; extension to more physical models, error control and uncertainty propagation in prediction, and compilers that map such irregular yet static scientific algorithms onto complex hardware topologies all remain open directions.