Maximum number of unknows

[valerio_19]

Hi folks!
What is the maximum dimension of a matrix can be handled by CULA (for example a LU-decomposition on a single precision complex matrix)?

Does it depend on the DRAM of the GPU?
Is it possible to perform matricial operations on matrix that cannot be stored in the GPU?

I'm a new in CULA and I have done some tests.

On my PC with poor GPU performances (NVIDIA QUADRO NVS 290 - 256 MByte) I have performed a lu decomposition with MKL-lapack and CULA Basic R11 . These are my results in seconds:

Unknowns T-lapack T-cula
1000 0 1
2000 1 3
3000 4 8
4000 8 18

The cpu is a Intel dual-core @ 3GHz.
The number of CPU threads has been set to 1.
The OS is windows XP 32 bit.

The performance difference is probably due to the view-card not performing.
However, if the number of unknowns increases, I get by the CULA routine an "insufficient memory error".
Is it correct? I hoped that CULA routines might solve larger problem.

Thanks,
Valerio

[kyle]

The NVS-290 is not a very good choice for running CUDA programs. Newer GPUs are about 32x times more powerful and have up to 24x memory. These GPUs are capable of solving matrices with over 20,000 unknowns in only a few seconds.

[valerio_19]

Yes, I imagined it, I wanted just try some simple CULA routines. However, I read old posts and I understand that the matrix dimension cannot be greater than the GPU memory.

Thanks,
Valerio

[kyle]

That is a current limitation. However, we are working "out-of-core" solvers that can bypass this restriction.