Machine for machine learning & simulation

We are considering machines for performing simulations such as machine learning and quantum chemical calculations.
Quantum chemical calculation assumes the use of VASP 6 and QuantumEspresso.
Since we want to assume GPU driving such as OpenACC, we want to install a GPU with a large video memory capacity after securing the specifications of the CPU and memory.
The assumed requirements are as follows.

・CPU: Latest generation as much as possible, high clock speed and large cache
・Memory: 1 GB or more per CPU core
・Storage: SSD x1, HDD of 1TB or more
・GPU: GPU with 20GB memory or more
・Power supply: 100V

Based on the contents of the consultation, we proposed the following specifications.

CPU Xeon Gold 6426Y (2.50GHz 16 cores) x2
memory 512GB ECC REG
storage 960GB SSD U.2
video NVIDIA T400 4GB MiniDisplayPort x3
network on board (10GBase-T x2)
Housing + power supply Full Tower Chassis + Redundant Power Supply 1000W
OS Ubuntu 22.04
GPU NVIDIA H100 80GB PCI-E

The CPU uses Xeon Scalable (Gold 6426Y x 2 sheets, total 32 cores), the memory is 512GB (1GB per core), and the GPU is NVIDIA H16 100GB.

Both VASP 6 and Quantum Espresso support double precision operations on GPGPU.In using these software, it can be said that there is an advantage to using NVIDIA H100 instead of Geforce.

The configuration of this case study is based on the conditions given by the customer.
Please feel free to contact us even if you are considering different conditions from what is posted.

 

 

■ Keywords

・What is VASP?
VASP (Vienna Ab initio Simulation Package) is material science software used for first-principles calculations.It is used for electronic structure calculation, molecular dynamics simulation, prediction of thermodynamic properties, etc.It is based on Density Functional Theory (DFT) and supports many types of systems such as solids, surfaces and molecules.

reference:VASP (VASP Software GmbH) *Jumps to an external site

 

・What is Quantum ESPRESSO?
Quantum ESPRESSO is open source software written in Fortran95 distributed free of charge under the GNU General Public License.First-principles electronic structure calculation program.It is a first-principles calculation code based on density functional theory.It can perform atom position and cell structure optimization, electronic state and band structure calculations, phonon vibrational analysis, and neutron scattering calculations. It operates on multiple platforms and supports parallelization by MPI and OpenMP.It is widely used for theoretical analysis of material properties and search for new materials, and is one of the useful tools in computational biology.

reference:Quantum ESPRESSO *Jumps to an external site

 

・What is OpenACC?
OpenACC is an API specification for easy parallel computation using GPUs. Defined for Fortran, C, and C++.Traditionally, parallel computing using GPUs (GPGPU) was difficult to program and complicated to code, but OpenACC makes it easy.

 

・What is NVIDIA H100?
The NVIDIA H100 is NVIDIA's next-generation data center GPU, designed for the AI ​​and high-performance computing (HPC) fields. Based on the Ampere architecture, high-performance arithmetic with various precisions such as FP32, FP64, and INT8 is possible. Equipped with 2GB of HBM40e memory, achieving a memory bandwidth of 3TB/s.It can efficiently process large models and datasets. A technology called Structural SIMD speeds up INT8 operations by up to 80 times. The calculation performance of FP64 is twice that of FP32, and the inference processing performance of INT2 is up to 8 times that of FP32.

reference:Nvidia H100 *Jumps to an external site