Relevant thread:
[url]https://devtalk.nvidia.com/default/topic/985037/gtx-1070-quot-gpu-has-fallen-off-the-bus-quot-running-3d-games-in-arch-linux-/[/url]
Under Linux, compile & run gpu-burn – note this will stress GPUs close to 100%, so make sure the power supply is enough to handle the full load and enough ventilation is present. If a GPU errors during this test it is more than likely the GPU itself.
If the fall off the bus results are reproducible at full load, gpu-burn program will more than likely trigger it within 5 or 10 minutes of running. Given your configuration of an external PCI-E device, it is quite possible that either the cabling, interface card, or even the PCI-E slot that the interface card is connected to could be bad, or even that you do not have enough power in the expansion slot for all the GPUs.
If power is not the issue, my advice would be to test each GPU separately physically in the machine to rule out the expansion solution.
FWIW, I had a different Xid error (32) that was caused (most likely) by a bad PLX chip or slot on an ASRock motherboard: [url]ASRock X99 WS-E memory compatiblity - ASRock Forums - Page 3
If you are able to test on Windows, Unigine benchmarks tend to trigger hangs as well. That can help debug if it is an O/S type issue or if it is an actual hardware issue.