`test_dot3d[8-2-64-64-64-32-32-float32-float32]` fails on B580, Linux

[pbchekin]

FAILED language/test_core.py::test_dot3d[8-2-64-64-64-32-32-float32-float32] - AssertionError: 
Not equal to tolerance rtol=0.01, atol=0.01

Mismatched elements: 510 / 32768 (1.56%)
Max absolute difference: 22.157192
Max relative difference: 5313.6147
 x: array([[[-15.017813,   3.083206,  -3.932423, ...,  -6.029019,
          -9.865822,  -9.993592],
        [  3.607818,  -7.83863 ,  18.861736, ...,  12.048253,...
 y: array([[[-15.016216,   3.082651,  -3.927522, ...,  -6.029953,
          -9.862488,  -9.997168],
        [  3.610554,  -7.831057,  18.854351, ...,  12.045836,...

https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14440691069/job/40514125850