NBLOCK_FOCK

NBLOCK_FOCK = [integer]

Default: LNONCOLLINEAR	= 64	CPU build
	= 32	GPU build (OpenACC/OpenMP offload))
	= `OMP_NUM_THREADS`	CPU build that is compiled with OpenMP support and threading is active (`OMP_NUM_THREADS` > 1)

Description: Sets the number of orbitals that are processed simultaneously when computing the action of the Fock potential.

Instead of computing the action of the Fock potential on one orbital at a time, up to NBLOCK_FOCK orbitals are gathered and processed at once. This enables the use of matrix-matrix operations rather than matrix-vector operations, which is beneficial for performance on modern hardware.

Tuning NBLOCK_FOCK can significantly affect both the performance and memory consumption of hybrid functional calculations. Especially on GPUs, NBLOCK_FOCK should be tuned carefully to achieve optimal performance.

Tip: On GPU architectures, the optimal value of NBLOCK_FOCK depends strongly on the specific hardware and the number of bands. It is recommended to experiment with values in the range 16-64.

Related tags and articles