EXESS

Official Resources

• Homepage: https://barcagrp.com/exess/
• Documentation: https://barcagrp.com/exess/
• Source Repository: Closed (Waitlist/Academic access)
• License: Academic/Commercial

Overview

EXESS (Extreme-scale Electronic Structure System) is a GPU-native quantum chemistry code designed for extreme-scale ab initio molecular dynamics (AIMD) capabilities. It won the 2024 ACM Gordon Bell Prize for its ability to perform MP2-level AIMD simulations on systems with thousands of atoms, leveraging novel algorithms optimized for GPU architectures (NVIDIA).

Scientific domain: High-performance quantum chemistry, Ab Initio Molecular Dynamics (AIMD)
Target user community: HPC users, researchers needing large-scale accurate dynamics (MP2)

Theoretical Methods

• Hartree-Fock (HF)
• Density Functional Theory (DFT)
• Second-order Møller-Plesset perturbation theory (MP2)
• Ab Initio Molecular Dynamics (AIMD)
• Resolution of Identity (RI) approximations
• GPU-accelerated algorithms

Capabilities (CRITICAL)

• GPU-native implementation (CUDA)
• Extreme scalability (Summit, Frontier, Aurora scales)
• Large-scale MP2 calculations (1000+ atoms)
• Long-timescale AIMD at MP2 level
• Energy conservation in dynamics
• High floating-point efficiency
• Massively parallel execution

Key Strengths

GPU Optimization:

• Built from scratch for GPUs
• High percent peak flop utilization
• Minimal CPU-GPU transfer
• optimized Tensor contractions
• Scalable to thousands of GPUs

MP2 Dynamics:

• Accurate electron correlation
• Dispersion inclusion via MP2
• Feasible for large biological systems
• Beyond DFT accuracy for dynamics

Extreme Scale:

• Linear scaling or low-prefactor algorithms
• Handles 1000-2000 atoms at MP2 level
• Gordon Bell Prize performance
• State-of-the-art HPC

Inputs & Outputs

• Input formats:
  • EXESS input format
  • PDB/XYZ coordinates
  • Basis set library inputs
• Output data types:
  • Energies (HF, MP2)
  • Forces
  • Trajectories (XYZ/DCD)
  • Restart files
  • Performance metrics

Interfaces & Ecosystem

• HPC Systems: Designed for OLCF/ALCF supercomputers (Summit, Frontier, Aurora)
• NVIDIA Integration: Optimized for A100/H100 GPUs using CUDA and cuBLAS
• Analysis Tools: Standard trajectory analysis (VMD, MDAnalysis)
• Input Generation: Minimal input scripts, compatible with standard formats
• Output Parsing: Standard text output, easily parsable performance logs

Advanced Features

MP2-AIMD:

• On-the-fly forces
• Conserved energy dynamics
• Solvated systems
• Chemical reactions in solution

Algorithms:

• Rank-reduced operations
• Mixed precision utilization
• Asynchronous task scheduling
• Distributed memory management

Performance Characteristics

• Speed: Orders of magnitude faster than CPU codes for MP2
• Accuracy: MP2/CBS limit capabilities
• System size: 1000+ atoms (MP2)
• Memory: GPU memory constrained (managed)
• Parallelization: Multi-node Multi-GPU (MPI+CUDA)

Computational Cost

• MP2: Conventionally O(N^5), EXESS optimized
• Dynamics: Feasible ps/ns scales
• Hardware: High-end GPU clusters required
• Efficiency: High FLOP/watt

Limitations & Known Constraints

• Availability: Not open source (Waitlist)
• Hardware: Requires NVIDIA GPUs
• Features: Focused on energy/forces (MP2), less property analysis
• Documentation: Limited public docs

Comparison with Other Codes

• vs CP2K: EXESS focuses on MP2, CP2K on DFT
• vs GAMESS: EXESS GPU-native, faster for large MP2
• vs TeraChem: Both GPU, EXESS targets HPC/MP2 scale
• vs Psi4: EXESS is HPC dynamics focused
• Unique strength: Large-scale MP2 dynamics on GPUs

Application Areas

Biochemistry:

• Enzyme Reactions: MP2 accuracy for reaction mechanisms in large enzymes
• Solvation Dynamics: Accurate description of solvation shells including dispersion
• Ligand Binding: Free energy calculations with correlated methods
• Conformational Ensembles: Sampling complex landscapes with high accuracy

Materials Science:

• Liquid Structures: Reliable radial distribution functions from MP2
• Interfacial Chemistry: Solid-liquid interfaces with accurate electronic structure
• Nanoparticles: Dynamics of metallic and semiconductor clusters
• Battery Electrolytes: Solvation structures and transport mechanisms

Best Practices

System Setup:

• Pre-equilibration: thorough equilibration with classical MD before switching to EXESS
• Basis Sets: Use standard correlation-consistent basis sets (cc-pVDZ/TZ)
• Geometry: Ensure clean starting structures to avoid large initial forces

Hardware Utilization:

• GPU Selection: Target A100 or H100 nodes for maximum efficiency
• Memory Management: Monitor GPU memory usage for large basis sets
• Scaling: Test scaling on small number of nodes before full production run

Simulation Parameters:

• Timestep: Use appropriate timestep (0.5-1.0 fs) for AIMD
• Thermostats: Standard thermostats (Nose-Hoover) available
• Restart frequency: Write restarts frequently due to HPC time limits

Community and Support

• Barca group (Australian National University)
• HPC centers (OLCF, etc.)
• Gordon Bell community
• Academic collaborations

Verification & Sources

Primary sources:

1. Homepage: https://barcagrp.com/exess/
2. Gordon Bell Prize 2024 announcements
3. Barca group publications (JCTC, etc.)

Confidence: VERIFIED

• Status: Active HPC code
• Recognition: Gordon Bell Prize
• Existence: Confirmed via ANU/OLCF