Perfomance index benchmarks

../_images/pw-perf-index.svg ../_images/lcao-perf-index.svg ../_images/fd-perf-index.svg

See gpaw/benchmark/performance_index.py.

Test systems

name

cores

IBZ

bands

a [Å]

b [Å]

c [Å]

A [°]

B [°]

C [°]

C60-0

24

1

148

16.6

16.9

17.0

90.0

90.0

90.0

H2-0

24

1

2

6.0

6.0

6.7

90.0

90.0

90.0

Mo60S120-1

56

2

940

9.5

32.6

32.6

90.0

90.0

90.0

Al96-2

40

10

177

8.6

11.5

28.4

90.0

90.0

60.0

C72-2

24

3

177

14.8

14.8

10.0

90.0

90.0

120.0

OPt24-2

40

16

238

5.7

5.7

31.5

90.0

90.0

60.0

CrSi2As4-2M

24

30

37

3.8

3.8

22.3

90.0

90.0

120.0

ErGe-2M

24

20

22

3.9

4.1

12.1

90.0

90.0

90.0

MnVS2-2M

24

20

32

3.6

4.0

14.7

90.0

90.0

90.0

PtLi2O6-2M

24

16

44

5.4

5.4

32.4

90.0

90.0

120.0

VI2-2M

24

12

21

4.1

4.1

13.1

90.0

90.0

120.0

V3Cl6-2N

24

12

53

6.3

6.3

12.7

90.0

90.0

60.0

Bi2Se3-3

24

12

30

4.1

4.1

28.6

90.0

90.0

60.0

C2-3

24

408

8

2.5

2.5

2.5

60.0

60.0

60.0

Ga2F4N4H10-3

40

42

43

4.9

5.9

20.1

90.0

90.0

90.0

LiC8-3

24

160

24

4.8

4.8

3.7

90.0

90.0

120.0

Ti2Br6-3

24

90

44

6.5

6.5

7.8

65.5

65.5

59.8

Fe8-3M

24

26

62

5.0

5.0

5.0

109.5

109.5

109.5

Fe8O8-3M

40

147

100

5.7

6.4

5.7

63.7

60.3

63.7

Mn2O2-3M

24

20

34

6.4

6.4

6.4

60.0

60.0

60.0

Parameters

  • PBE

  • k-point density: 5.0 Å

  • 14 electron potential for Cr

    • PW: 800 eV plane-wave cutoff

    • LCAO: 0.15 Å grid-spacing and dzp basis sets

    • FD: 0.15 Å grid-spacing

  • default parameters for everything else

PW-mode performance index

The total time for one material is the sum of two steps (\(t=\Delta t_1 + \Delta t_2\)):

  1. time for for a complete SCF calculation

  2. time for second SCF calculation after a small displacement of positions (or cell)

gpaw.benchmark.performance_index.score(data)[source]

GPAW’s PW-index (or score).

With \(N\) materials and times for completion of each material \(t_i\), we get this index (normalized to 100 for the first run with reference times \(t_i^0\)):

\[\frac{100 α}{N} \sum^{N}_{i=1} \frac{t^{0}_{i}}{t_{i}} \cdot\]

The rescaling factor α is used for rescaling the index when new materials are added or hardware is updated.

../_images/score.png

Results

Figure shows:

  • \(t_i^0 / t_i\)

  • \(\Delta t_2 / t\)

  • Memory usage per core

../_images/benchmark.png

PW-mode results for latest version:

name

dt1 [sec]

iter1

dt2 [sec]

iter2

memory [Gbytes]

C60-0

94.9

13

74.4

10

1.07

H2-0

1.6

9

1.4

8

0.26

Mo60S120-1

2207.8

19

1982.9

17

3.43

Al96-2

567.9

19

330.8

11

1.53

C72-2

200.4

14

148.8

10

0.83

OPt24-2

408.7

29

419.5

30

0.86

CrSi2As4-2M

60.9

17

45.9

13

0.37

ErGe-2M

44.8

34

28.1

22

0.31

MnVS2-2M

36.9

24

17.8

12

0.29

PtLi2O6-2M

304.7

39

288.3

36

0.58

VI2-2M

24.7

17

10.0

7

0.44

V3Cl6-2N

1188.1

20

1129.9

19

0.78

Bi2Se3-3

50.9

32

10.2

6

0.35

C2-3

5.0

9

1.5

3

0.21

Ga2F4N4H10-3

60.1

16

15.8

4

0.62

LiC8-3

17.1

12

13.8

11

0.28

Ti2Br6-3

112.9

16

53.5

8

0.45

Fe8-3M

72.3

26

24.0

9

0.26

Fe8O8-3M

258.0

26

58.1

6

0.68

Mn2O2-3M

230.9

111

22.5

10

0.30

History

2025, July

  • Initial run with 14 systems (score set to 100.0).

  • Niflheim (xeon24el8, xeon40el8_clx, xeon56).

  • Easy-build foss-2025b toolchain (Python-3.13.5, Numpy-2.3.2, Scipy-1.16.1, Libxc-7.0).

2025, November

  • Added three more systems (MnVS2-2M, PtLi2O6-2M, V3Cl6-2N).

  • Switched to “New GPAW”.

2025, November 26

  • Changed initial magnetic moments for MnVS2-2M from [2, 2, 0, 0] to [2, -2, 0, 0] and rescaled timings. See MR: !3032.

2025, December

  • Added LCAO and FD results to the graphs.

2026, January

  • Added two more systems (ErGe-2M, Fe8O8-3M).

2026, April

  • Added one more system: Mn2O2-3M.

  • Modified all the initial magnetic moments, s.t. the correct ground states are found.

  • Modified expected timings, s.t. the PW score remains unchanged.