Mova runs in two modes. CPU only is the reference: exact, and what the school run shows. GPU enhanced (MOVA_EXEC=gpu) reads the lessons and the tests with the GPU parser instead of the CPU one, an approximant; the student itself, its memory and its choices, is the same code. Here both modes take the same courses, test by test.
75.1%the same choice · 37275 of 49635 tests
+193tests right on the GPU · 18882 against 18689 on the CPU (38.04% against 37.65%)
18 / 6courses gained / lost · on the GPU
1.32× slowerthe GPU against the CPU · 1499 s on the CPU, 1975 s on the GPU, three courses at a time
The two parsers draw some trees differently, so the student reads some sentences into other frames and chooses another option there. «Same choice» counts the tests where both modes chose the same option, right or wrong. The GPU run is for training and bulk measurements; scores on the site come from the CPU. CPU run: 2026-10-22, commit 249ff72f1; GPU run: 2026-10-22, commit 249ff72f1.
Mova
