I’m interested in the maximum number of CUDA-devices that are supported by the CUDA-driver. I have never heard about a software limitation, but I’m not sure if e.g. 32 graphic cards in one system would be supported by the driver.
The reason why I’m asking is, that our institute of computer science is going to buy a new computer with a lot of graphic cards in it. We already found a mainboard which has enough pcie slots for 32 cards but the question is, are they all supported by the driver at once.
I hope someone could answer my question, thanks in advance!
I think it is safer to go with a builder like HP for example. They have a lots of experience with hpc co mputers, though they are expensive. I am only mentioning the, because 32 cards raise a few problems such a power consumption 32x225 W and the heat problem as well.
of course we won’t configure the system on our own, some builder will be contacted. But the problem is, that first of all some money is needed from third-party funds and in order to get this a request is needed. As you can maybe imagine it would be a little bit embarrissing if the system we want to purchase, can’t actually be configured. Which means we won’t get the money anyway.
This is why I’m first of all interested in how many cards could be used on a single board with CUDA, power consumptions and heat problem is yet another problem.
We’ve run into BIOS issues with quite a few machines working with more than 4 GPUs. We’ve even had a few systems where BIOS issues prevented connecting more than 2!! Lesson: Make sure that when you contract a third party to build the hardware, that you include the necessary BIOS support as a requirement.