MDWF performance comparison - callat-qcd/chroma GitHub Wiki

testing the performance of the new support through Chroma of QUDA's various MDWF solvers.

We report on performance test using the a15m135XL ensemble with 2 Summit nodes with -geom 1 1 3 4.

The time is reported in seconds and the performance is reported in GFlops.

The times reported here do not account for the need for a tighter tolerance with the EES solver, but are from a fixed request of 1.e-7.

OOA Prec EES (`1e-7`) EES (`5e-8`)
Iter. time Perf. Chroma resid Iter. time Perf. Chroma resid Iter. time Perf. Chroma resid
11772 674.6 20098 1.75e-06 8409 494.8 29074 5.73e-06 9764 551.4 30295 1.68e-06
12783 703.5 20924 1.55e-06 9500 553.9 29340 3.32e-06 10487 618.5 29007 1.85e-06
11897 649.8 21084 1.74e-06 8887 537.9 28263 3.64e-06 9505 581.4 27970 2.00e-06
12386 675.4 21118 1.88e-06 9814 584.4 28730 2.94e-06
11466 625.8 21099 2.26e-06 8588 507.3 28961 3.88e-06
12608 686.0 21163 1.73e-06 9625 564.9 29152 3.93e-06
total QUDA 4015 3243
total solve 4656.3 3739.1
⚠️ **GitHub.com Fallback** ⚠️