maximum number of GPUs supported?

Accelerated Computing CUDA CUDA Programming and Performance
1

How many GPUs can be driven by the latest NVIDIA Linux driver (256.44)? I’ve been trying to set up a system connected to 12 GPUs (in 3 Tesla S2050 boxes), but each time I attempt to initialize CUDA (or run the nvidia-smi command, for that matter), I observe the following error:

NVIDIA: could not open the device file /dev/nvidia8 (Input/output error).

(Each S2050 contains 4 GPUs; hence, the above error is occurring when the system tries to initialize the 9th GPU.)
Concurrently, the following error appears in the system’s dmesg output:

NVRM: RmInitAdapter failed! (0x54:0x20:1701)
NVRM: rm_init_adapter(8) failed

I’m pretty sure the hardware is working fine; when I disconnect any 4 GPUs, initialization proceeds without any issues.

2

It depends on the system’s BIOS. The driver has been tested with up to 16 GPUs, but often times system BIOSes will get very upset with having that much BAR space and won’t behave correctly.

3

It depends on the system’s BIOS. The driver has been tested with up to 16 GPUs, but often times system BIOSes will get very upset with having that much BAR space and won’t behave correctly.