GPU Acceleration (GPU4PySCF)#
Modules: gpu4pyscf
Introduction#
Modern GPUs accelerate quantum chemistry calculation significantly, but also have an advantage in cost saving [1]. Some of basic PySCF modules, such as SCF and DFT, are accelerated with GPU via a plugin package GPU4PySCF (See the end of this page for the supported functionalities). For the density fitting scheme, GPU4PySCF on A10080G can be 1000x faster than PySCF on singlecore CPU. The speedup of direct SCF scheme is relatively low.
Installation#
The binary package of GPU4PySCF is released based on the CUDA version.
CUDA version 
GPU4PySCF 
cuTensor 
CUDA 11.x 


CUDA 12.x 


Usage of GPU4PySCF#
GPU4PySCF APIs are designed to be compatible with PySCF. When supported, highlevel functions and objects are named the same as PySCF. But, GPU4PySCF classes do not directly inherit from PySCF class.
PySCF objects and GPU4PySCF objects can be converted into each other by to_gpu()
and to_cpu()
. In the conversion, the numpy arrays will be converted into cupy array. And the functions will be omitted if they are not supported with GPU acceleration.
One can use the two modes to accelerate the calculations: directly use GPU4PySCF object:
import pyscf
from gpu4pyscf.dft import rks
atom ='''
O 0.0000000000 0.0000000000 0.1174000000
H 0.7570000000 0.0000000000 0.4696000000
H 0.7570000000 0.0000000000 0.4696000000
'''
mol = pyscf.M(atom=atom, basis='def2tzvpp')
mf = rks.RKS(mol, xc='LDA').density_fit()
e_dft = mf.kernel() # compute total energy
print(f"total energy = {e_dft}")
g = mf.nuc_grad_method()
g_dft = g.kernel() # compute analytical gradient
h = mf.Hessian()
h_dft = h.kernel() # compute analytical Hessian
Alternatively, one can convert PySCF object to the corresponding GPU4PySCF object with to_gpu()
since PySCF 2.5.0
import pyscf
from pyscf.dft import rks
atom ='''
O 0.0000000000 0.0000000000 0.1174000000
H 0.7570000000 0.0000000000 0.4696000000
H 0.7570000000 0.0000000000 0.4696000000
'''
mol = pyscf.M(atom=atom, basis='def2tzvpp')
mf = rks.RKS(mol, xc='LDA').density_fit().to_gpu() # move PySCF object to GPU4PySCF object
e_dft = mf.kernel() # compute total energy
When the GPU task is done, the GPU4PySCF object can be converted into the corresponding PySCF object via mf.to_cpu()
.
Then, more sophisticated methods in PySCF can apply. One can also convert the individual CuPy array to numpy array with `Cupy APIs`_.
Functionalities supported by GPU4PySCF#
Method 
SCF 
Gradient 
Hessian 
direct SCF 
O 
GPU 
CPU 
density fitting 
O 
O 
O 
LDA 
O 
O 
O 
GGA 
O 
O 
O 
mGGA 
O 
O 
O 
hybrid 
O 
O 
O 
unrestricted 
O 
O 
O 
PCM solvent 
GPU 
GPU 
FD 
SMD solvent 
GPU 
GPU 
FD 
dispersion correction 
CPU* 
CPU* 
FD 
nonlocal correlation 
O 
O 
NA 
ECP 
CPU 
CPU 
CPU 
MP2 
GPU 
CPU 
CPU 
CCSD 
GPU 
CPU 
NA 