You see IOMMU faults if the IOTLB is not properly invalidated.
You see interrupt steering issues if the interrupt remapping table is not invalidated properly.
In the /var/log/vmkernel.log file on the ESXi host, you see entries similar to:
This issue occurs because the ESXi host does not give enough time for COMPLETION_WAIT commands to complete.
Note: This is limited to AMD IOMMU only. This article is not relevant to Intel IOMMU.
This is a known issue affecting VMware ESXi 6.0.x, 6.5.x, and 7.0.2. Currently, there is no resolution or workaround for ESXI 6.0 and 6.5.
This is fixed for ESXi 6.7.
For 7.0.2, there is a software issue which can trigger the issue with low probability if system uses DMA map/unmap heavily. This has been been fixed in 7.0.3 release.