Installing Tesla P100 on Ubuntu 16.04 Server with a 1060GTX

Accelerated Computing CUDA CUDA Setup and Installation
1

I’m having trouble installing a a P100 in this system. I tried to manually install the driver using sudo dpkg -i nvnameofdriver.deb (downloaded 375 ubuntu driver from the nvdia site). It indicates OK after the install, but nvidia-smi only detects the 1060.

Is there anything I am missing for installing the P100? Is there a different way of checking if the non GPU card is installed?

Does it even work in a “multi-GPU” mode? I’m working on Tensorflow and want to shift my dev to the P100 now. I actually have 2 cards and this is a preliminary test system built on a z170/i7 system. Eventually I will spec out a full system so I can have 2 P100 cards installed, but getting stalled on just 1 card install is a bit rough.

2

Maybe your P100 is overheating. It’s not really designed to be installed in a custom-built system, but rather in an OEM certified system designed to accept the P100.

There may also be incompatibilities with the system BIOS that prevent the P100 from working correctly.

3

It’s not overheating, its not even warm to the touch. From a BIOS perspective, it is just a PCI-E card. I don’t recall anything that would discriminate the P100 from any other GPU from a BIOS perspective. The newer Tesla cards are supposed to be more open in their design than previous workstation cards.

4

what is the output of:

dmesg | grep NVRM

and

sudo lspci -vvv |grep -A 15 3D

5

That determination may or may not be useful depending on where you touch the card.

Not sure what is the point of that. All PCIE cards are the same? No, they are not.

Here is an example of an incompatibility between a system BIOS and a PCIE GPU that has a large BAR aperture:

https://devtalk.nvidia.com/default/topic/790759/cannot-install-driver-for-nvidia-tesla-k40-cards-on-fedora-20/

The net effect is that the system BIOS fails to map PCIE resources needed by the card, and as a result the card is not usable. P100 has a similarly large BAR aperture.

6

I ended up getting to the BAR error after updating the BIOS and freeing up the PCI-E lanes. I put in an order for a Dell R730 as its on the qualified list. I think there’s a lot of potential in this on desktop research as well if we can get some key players involved. I guess I’ll be pushing the matter in May when I visit the nv team there.