Just lost. If you run the code with just 32 processors, it runs super-fast. All pinned memory, no 4090 board and just the 32 processors, also runs super-fast. All pinned memory while using the 4090 board along with the 32 processors, significantly slower. Why? Any help would be appreciated. Completely out of ideas. Been working on this problem for weeks. Thanks in advance.
It is impossible to even comment intelligently without knowing your hardware configuration, what kind of software you are running, and how it uses the GPU.
GPUs are not magic bullets: Not every use-case is suitable for GPU acceleration, and even if the workload is amenable to GPU acceleration in principle there are ways for programmers to shoot themselves in the foot. There are also ways to hamstring a GPU-accelerated system via poor hardware configuration (example: a common mistake is an undersized system memory).
I would start with profiling. Start with Nsight Systems, then drill down on the GPU portion with Nsight Compute.
BTW, when you say “32 processors” I assume you mean a single CPU with 32 cores?
Agree with njuffa - the question needs more details,
you probably were exhausted from trying things.
Take some deep breaths and edit your question to give us enough information to know, what you are actually doing.
So the 32 (CPU?) processors (cores? threads?) are still running your code? Or you ported to GPU?
What do you mean by no 4090 board vs. using the board? no 4090 board = it not being mounted in the system, it sitting idly in the system? using the board = having a screen connected, it running different software, it running your code?