Hardware Requirements

Software Versions Used in This Tutorial

The results presented in Sahrawat et al. (2024) were produced with the following software versions. Using different versions may require minor adjustments to input flags or file formats.

Software	Version	Notes
AMBER	22 (SANDER 2022)	Classical MD with pmemd.cuda; QM/MM production and SMD with sander
TeraChem	v1.96H-beta (build 2023-04-07)	Development build; compiled against CUDA 12.1; supports SM 5.0–9.0 (Turing, Ampere, Ada)
NBO	7.0 (Linux x64)	Binary distribution; called via TeraChem at each QM/MM SMD step

Note

TeraChem v1.96H-beta is a development release from PetaChem. If you are using a stable release or a newer version, the input syntax and available keywords should remain compatible, but confirm with the TeraChem release notes if you encounter unexpected behaviour.

What You Need to Run This Tutorial

This tutorial couples three distinct computational tools — AMBER, TeraChem, and NBO — and each has different hardware demands. The table below gives a quick overview; detailed guidance for each component follows.

Software	Step in Tutorial	Hardware	Notes
pmemd.cuda	Classical MD (Steps 1–3)	Nvidia GPU	Any modern Nvidia GPU; CUDA-enabled
sander	QM/MM minimisation and production MD	CPU (multi-core)	AMBER’s QM/MM interface does not use the GPU
TeraChem	QM calculations within QM/MM	Nvidia GPU (CUDA)	CUDA core count is the key metric, not vRAM
NBO	NBO analysis alongside QM/MM SMD	CPU (multi-core)	Runs on CPU only; scales well with core count

TeraChem: GPU is Essential, CUDA Cores are the Key

TeraChem performs all its electronic structure calculations directly on the GPU. It is written entirely in CUDA and achieves substantial speedups over CPU-based QM packages precisely because it maps the linear-algebra operations of DFT onto the massive parallelism of a GPU’s shader cores.

What matters most: CUDA core count, not vRAM.

The QM region in a typical enzyme simulation (50–150 atoms with a 6-31G* basis) fits comfortably within 4–8 GB of GPU memory. What limits throughput is floating-point throughput, i.e., the number of CUDA cores running in parallel. A high-end gaming GPU with a large core count will therefore outperform a workstation GPU that has more vRAM but fewer cores.

Best price-to-performance GPUs for TeraChem

Consumer-grade Nvidia GPUs offer an excellent price-to-performance ratio for TeraChem. Our recommended choices (in order of increasing performance):

• RTX 3090 — 10,496 CUDA cores, 24 GB GDDR6X. A strong performer for QM/MM and widely available second-hand.

• RTX 4090 — 16,384 CUDA cores, 24 GB GDDR6X. Currently one of the best single-GPU options for TeraChem.

• RTX 5090 — 21,760 CUDA cores, 32 GB GDDR7. The latest generation; best single-GPU performance available as of 2025.

For multi-GPU runs (ngpus = 2 in the TeraChem template), two RTX 4090s or two RTX 3090s are a very practical and cost-effective setup. The demo version of TeraChem supports up to two GPUs and a maximum runtime of 15 minutes per job, which is sufficient for the QM/MM MD steps in this tutorial.

Note

TeraChem requires an Nvidia GPU with CUDA compute capability ≥ 3.5. AMD and Intel GPUs are not supported. Always install the CUDA toolkit version that matches your TeraChem build.

AMBER pmemd.cuda: GPU-Accelerated Classical MD

The classical MD steps (minimisation, thermalisation, NPT equilibration in System Equilibration and Minimisation) use pmemd.cuda, AMBER’s GPU-accelerated production engine. For these steps the GPU is used primarily for non-bonded force evaluation, and any modern Nvidia GPU (Turing architecture or newer) will provide a substantial speedup over CPU-only pmemd.

There is no strict minimum vRAM for the system sizes typical in this tutorial (~70,000 atoms with solvent), but 8 GB or more is comfortable. The same RTX 3090/4090/5090 GPUs recommended for TeraChem work equally well here.

Important

pmemd.cuda and TeraChem cannot share the same GPU simultaneously, as TeraChem claims the full GPU memory for its QM calculations. In practice this is not an issue because the pmemd.cuda classical equilibration steps complete before the QM/MM production runs begin. On a multi-GPU node you can run pmemd.cuda on one GPU while TeraChem runs on another.

AMBER sander: QM/MM Runs on CPU

The QM/MM production runs and SMD simulations (QM/MM Production Run with TeraChem, QM/MM Steered Molecular Dynamics (SMD)) use AMBER’s sander engine, not pmemd.cuda. This is because the QM/MM interface in AMBER — the file-based protocol that hands coordinates to TeraChem and reads back energies and gradients — is implemented only in sander. The MM part of the force evaluation in sander runs on CPU.

In practice, sander is not the computational bottleneck: TeraChem on the GPU handles the expensive QM step, and sander simply manages the I/O and MM calculation between QM calls. A modern multi-core workstation CPU (8–32 cores) is more than sufficient for the sander side.

NBO: CPU-Only, Scales with Core Count

NBO (Natural Bond Orbital) analysis runs entirely on CPU. At each step of the NBO-coupled QM/MM SMD simulations (Natural Bond Orbital (NBO) Analysis), TeraChem writes a .47 wavefunction file and calls the NBO binary, which processes the orbital information and returns the results to TeraChem before the next MD step proceeds.

NBO7 parallelises well across CPU cores; providing 8–16 dedicated CPU cores will keep the NBO step from becoming a bottleneck in the overall QM/MM SMD pipeline. Ensure the NBO executable path is correctly set in your environment before running these simulations:

export NBOEXE="$your_nbo_install_dir/bin/nbo7.i4.exe"

Summary: a Practical Single-Workstation Setup

The simulations in this tutorial were carried out on a workstation with two consumer-grade Nvidia GPUs. The following configuration is representative of what you need:

• GPU: 2 × Nvidia RTX 3090 (or RTX 4090 / 5090 for better performance)

• CPU: 16–32 cores (Intel or AMD) for sander MM evaluation and NBO analysis

• RAM: 64–128 GB system RAM (TeraChem and NBO both make heavy use of system memory)

• Storage: Fast NVMe SSD recommended — QM/MM SMD runs generate large numbers of small files (TeraChem scratch, NBO .47 files, AMBER restart files) and I/O speed matters

Cloud-based GPU instances (e.g., AWS p3/p4, Google Cloud A100 nodes) can also be used, but consumer-grade gaming GPUs remain the most cost-effective option for research groups running TeraChem at this scale.