Known issues

NVIDIA

• No issues at the moment

AMD

• HIPBLAS bug affecting DGEMMs on AMD systems.

This bug is particularly nasty, it presents itself for very specific matrix sizes. It commonly appears when running multi-GPU runs of RI-HF and RI-MP2 gradient calculations. Any calculation type that uses a DGEMM has a potential to be affected.

The use of rtat can make this happen more often, since it tests diverse combinations for performance. If disabling rtat does not work for you, please file a bug report with us.

• DFT code produces incorrect answers on AMD GPUs (as of v5.0.0 of EXESS)

• Out of resources issue. If a similar printout as the one below happens, edit the amount of max_gpu_memory_mb you are using. See max_gpu_memory_mb. This issue seems to also pop up more the more GPUs you ask for when running “small” systems, (20-40 atoms). This bug seems to be caused by the use of four center kernels. Use RI in fock_build_type to circumvent this issue.

:0:rocdevice.cpp
:2688: 1214497773164 us: [pid:853101 tid:0x14e50c57d700]
Callback: Queue 0x14dbb9800000 Aborting with error :
HSA_STATUS_ERROR_OUT_OF_RESOURCES: