Mixed-Precision Ab Initio Tensor Network State Methods Adapted for NVIDIA Blackwell Technology via Emulated FP64...

Area of Science:

• Quantum Chemistry
• Computational Physics
• Materials Science

Background:

• Accurate electronic structure calculations are crucial for understanding chemical reactions and material properties.
• High-precision arithmetic (FP64) is computationally expensive, limiting the scale of simulations.
• Developing efficient computational methods is essential for advancing quantum chemistry.

Purpose of the Study:

• To evaluate the performance of mixed-precision spin-adapted ab initio Density Matrix Renormalization Group (DMRG) calculations.
• To demonstrate the feasibility of emulating FP64 arithmetic using fixed-point resources for correlated calculations.
• To assess the accuracy and potential of new hardware technologies for large-scale electronic structure computations.

Main Methods:

• Utilized the Ozaki scheme for emulating FP64 arithmetic with fixed-point compute resources.
• Employed mixed-precision arithmetic in spin-adapted DMRG calculations.
• Performed calculations on benchmark systems and active sites of FeMoco and cytochrome P450 enzymes with large complete active space (CAS) sizes.

Main Results:

• Achieved milli-Hartree accuracy in electronic structure calculations using mixed-precision arithmetic.
• Demonstrated that DMRG's variational nature is suitable for benchmarking hardware and numerical libraries.
• Presented detailed numerical error analysis and performance assessments for DMRG subcomponents.

Conclusions:

• Mixed-precision emulation of FP64 arithmetic using fixed-point resources can achieve chemical accuracy in quantum chemistry.
• This approach paves the way for utilizing advanced hardware, like Blackwell technology, in tensor network state calculations.
• Opens new avenues for research in materials science and other fields requiring large-scale correlated electronic structure simulations.