note	kmesh	config	wall_s	formula	n_atoms	hardware
Stage 1: production TB2J before the rewrite	8x8x6	prod-a100	240.7	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	a100-cusolver	254.9	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	a100-mpi8	185.8	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	cpu8-mpi8	288	FePt	2	8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	cpu8-mpi8-kpar1	430.2	FePt	2	8 CPU
FePt after the kernel rewrite and H(R) cache	8x8x6	cpu8-mpi8	57.6	FePt	2	8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	prod-a100	560.7	FePt	16	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	a100-cusolver	283.8	FePt	16	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	cpu8-mpi8	340.5	FePt	16	8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	cpu16-mpi16	320.9	FePt	16	16 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	5x5x6	prod-a100	417.3	GdCo5	6	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	5x5x6	a100-cusolver	309.9	GdCo5	6	A100 + 8 CPU

Sign in to see the full dataset

Sep 29

Ouro DFT MAE route benchmark: hardware, process layout, TB2J kernel

Data Views Docs Logs

Overview

33

Attribution

Connections

1

Referenced by

Ouro DFT's MAE route: an idle A100, a slow TB2J loop, and a band cut that made results depend on the cell
MAE jobs now run on CPUs, 2–4× faster and 6–12× cheaper. A TB2J band cutoff that biased FePt by 10% and made the MAE depend on the cell is fixed; results now match self-consistent SOC to within 5 µeV. MAE also respects ferrimagnetic seeds and +U.

Attribution

CC-BY-4.0 license

Wall time, per-stage timing, GPU utilisation and cost for the /dft/magnetic/mae (TB2J) route. Cases are L1₀ FePt (2 and 16 atoms) and ferrimagnetic GdCo₅. Runs compare A100 and CPU workers, the ELPA, genelpa and cusolver solvers, and MPI process layouts, before and after the TB2J kernel rewrite. Every run in a case must give the same MAE. MAE values here still use TB2J's former 5.1 eV band cut; the accuracy dataset has the corrected values. Costs use Modal list prices.

note	kmesh	config	wall_s	formula	n_atoms	hardware
Stage 1: production TB2J before the rewrite	8x8x6	prod-a100	240.7	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	a100-cusolver	254.9	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	a100-mpi8	185.8	FePt	2	A100 + 8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	cpu8-mpi8	288	FePt	2	8 CPU
Stage 1: production TB2J before the rewrite	8x8x6	cpu8-mpi8-kpar1	430.2	FePt	2	8 CPU
FePt after the kernel rewrite and H(R) cache	8x8x6	cpu8-mpi8	57.6	FePt	2	8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	prod-a100	560.7	FePt	16	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	a100-cusolver	283.8	FePt	16	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	cpu8-mpi8	340.5	FePt	16	8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	4x4x3	cpu16-mpi16	320.9	FePt	16	16 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	5x5x6	prod-a100	417.3	GdCo5	6	A100 + 8 CPU
Stage 2: bigger cells, final TB2J kernel; MAE still with the 5.1 eV band cut	5x5x6	a100-cusolver	309.9	GdCo5	6	A100 + 8 CPU

Sign in to see the full dataset

3

Stats

Rows: 14
Columns: 20
Size: 32 KB
Downloads: 0

Join to comment

Connections

1

Referenced by

Ouro DFT's MAE route: an idle A100, a slow TB2J loop, and a band cut that made results depend on the cell
MAE jobs now run on CPUs, 2–4× faster and 6–12× cheaper. A TB2J band cutoff that biased FePt by 10% and made the MAE depend on the cell is fixed; results now match self-consistent SOC to within 5 µeV. MAE also respects ferrimagnetic seeds and +U.