Plotting stalls and crashes with (2) GPU on Multi Socket system #2
Description
So far I am able to plot fine on one GPU. I'm was curious about dual GPU plotting so I'm looking into it.
The problem right now is the plotting crashes any time I run in with 2 GPU enabled.
Dell T7610
Windows 10 Pro
Dual Intel E5-2670v2
Dual 3060 Ti FE
512MB 1866Mhz DDR3
Maybe 512 just simply isn't enought RAM for 2 GPUs?
Also I tried with NUMA enabled and disabled with the same results
Looking at MadMax documentation on multi socket systems and multiple GPUs it looks like NUMA should be enabled and then each GPU assigned to its respective CPU and looks like there is some commands to do so in the plotter code but to me it doesn's look like Plot Manager has this option or inserts the commands to do so.
**It's recommended to run one GPU per CPU in case of multi-socket machines, while making sure to match each GPU with the correct CPU that's directly connected, and restricting memory allocations to local RAM.
Example:
numactl -N 0 -m 0 ./cuda_plot_k32 -g 0 ...
numactl -N 1 -m 1 ./cuda_plot_k32 -g 1 ...**