Speeding up Quantum Espresso by 36x using Claude Fable 5.2

Discovered Materials used Claude Fable 5.2 and a set of test suites to patch Quantum ESPRESSO, speeding DFPT phonon calculations by 36x and hybrid-functional DFT by 9x, according to discoveredmaterials.com's report. The open-source patch is the GitHub repository discovered-materials/dm_qe. Quantum ESPRESSO, roughly 600,000 lines of Fortran, is the most widely used open-source DFT code, and DFT is consistently among the largest consumers of time on national supercomputers.

This piece covers both speedups, the GPU kernel changes, the benchmark timings, and a nitrogen-vacancy run in diamond. It is for researchers who use Quantum ESPRESSO on semiconductors, batteries, catalysts, solar absorbers, aerospace alloys, or drug-discovery energetics, and who skip hybrids or phonons because stock runs are too slow.

01Band gaps, defects, and phonons

Density functional theory is the workhorse of computational material science and the most widely used method for simulating matter at the atomic scale. It is used for semiconductors, battery cathodes and solid electrolytes, catalysts for fertilizer and hydrogen production, solar absorbers, and aerospace alloys. In drug discovery the same method supplies binding energies, charge distributions, and reaction pathways that classical force fields cannot capture.

Conventional DFT is comparatively cheap, but it systematically underestimates band gaps and misplaces defect levels, which matters enormously for semiconductors. Hybrid functionals account more accurately for the Pauli exclusion principle and offer a promising remedy. They are much slower, which limits them to certain smaller toy systems rather than realistic large scale simulations.

Quantum ESPRESSO also computes how atoms vibrate, through density functional perturbation theory. Those phonons support crystal stability, heat capacity, thermal conductivity, dielectric response, and conventional superconductivity. DFPT costs an order of magnitude more than DFT, phonon data cover only a subset of the Materials Project, and screens often skip dynamic stability checks.

02Kernel rewrites and hidden tests

Discovered Materials gave Claude Fable 5.2 two tasks: speed up the phonon code, and speed up the hybrid-functional algorithms on GPUs. Fable 5.2 rewrites Quantum ESPRESSO kernels for the GPU with mixed precision, batching, and symmetry, then iterates on every failure. Timings, failures, and regressions go back into the next patch.

The loop checks standard benchmark suites and hidden test sets that span representative systems, so a speedup is not overfit to publicly known tests. The Discovered Materials test suite holds reference results from stock Quantum ESPRESSO and scores speed and correctness. Only changes that are faster and match those references are released, and any leftover mismatches fall within the measurement errors of experimental characterization.

03Hybrid functionals nine times faster

On the hybrid task, Fable 5.2 sped Quantum ESPRESSO up by roughly 9x as a geometric mean and, without being asked, added a multi-GPU version that spreads the work across devices and larger cells. MgO PBE0 vc-relax, Si-216 PBE0, 64 H2O PBE0 MD, and C60 B3LYP cannot run on the stock four-GPU path. Optimized four-GPU timings for those four are 6.1, 26.5, 118.7, and 22.7.

The orbital buffer stays on the GPU between applications of the exchange operator. Stock Quantum ESPRESSO copied it from the host on every application, which created large PCIe traffic latencies. Band pairs, handled one at a time with many small GPU launches, are grouped into blocks sized to the GPU's L2 cache and run with batched FFTs and a few fused kernels.

Building the real-space orbitals at each exchange initialization, FFTs and symmetry rotations included, now runs on the GPU instead of the CPU. That removes a serial bottleneck and a large host-to-GPU transfer, and the exchange contribution to the stress uses the same batched machinery. Small subspace diagonalizations in the Davidson solver stay on the CPU, because launch and synchronization overhead outweighs the computation at that size.

04Phonon runs thirty-six times faster

Fable 5.2's DFPT phonon code is roughly 36x faster than stock Quantum ESPRESSO on the same GPU, a geometric mean across 13 cases. Against stock CPU the geometric mean is 14x, on the 12 cases that have a CPU reference. Wall-clock time for the benchmark suite fell from 3.3 hours to under 4 minutes.

The four changes preserve backward compatibility with existing Quantum ESPRESSO inputs and workflows. Mixed precision runs the bulk of the linear-response solves in lower precision where accuracy permits, and promotes to double precision only where it matters. Crystal symmetry avoids recomputing equivalent perturbations, and a better preconditioner improves convergence so fewer iterations are needed per perturbation.

Setup work that previously ran on the CPU now runs on the GPU, removing a serial bottleneck and host-device transfers. Discovered Materials says the changes require electronic structure physics, GPU kernels, and Fortran, "a combination of skills that is rare to find amongst experts in the field." The largest versus-GPU factor in the table is 108x, on diamond PAW with an 11×11×11 k-mesh and a low-symmetry q.

05NV centers on four H100s

Discovered Materials put both speedups on the nitrogen-vacancy center in diamond, an NV-defect quantum sensor and a leading candidate for room-temperature qubits. Defect level positions need hybrid functionals, and electron-phonon coupling needs DFPT. The patched code that passed the test suite is discovered-materials/dm_qe, and the phonon path still accepts existing Quantum ESPRESSO inputs.

In the benchmark, the phonon calculation for a negatively charged NV center in a 63-atom diamond supercell went from 77 seconds on stock GPU Quantum ESPRESSO to 6 seconds, called 13x faster. The row nv64_nc_lsda_gamma, an LSDA run of 4 modes, lists stock CPU 178.9, stock GPU 77.2, and optimized 6.1, or 12.7x versus the stock GPU and 29x versus the stock CPU. With the multi-GPU hybrid implementation, the same system can be treated at hybrid accuracy on four H100s.

Work that took several days now finishes in a few hours and can screen hundreds of defect-host combinations. Perdew's Jacob's ladder runs from (semi-)local LDA and GGA through meta-GGAs and hybrids to double hybrids and RPA, each rung closer to the heaven of chemical accuracy and an order of magnitude or more costlier. Discovered Materials calls that a grand challenge, after decades of picking a rung and living with its errors, and is optimistic LLMs can rapidly collapse it and bring hybrid, double-hybrid, and RPA-level accuracy to today's GGA speed.