RAM issue or GPU issue?

My task is to perform iterative 3D MRI image reconstruction through GPU. The matrix size (data) would be 384 x 384 x 384 x 30 x 5 x 2 with single precision, complex numbers. That is usual matrix size, and would be extended larger (512 x 512 x 512 x 30 x 30 x 10). For this, 'gpuNUFFT' function in MATLAB ( https://github.com/andyschwarzl/gpuNUFFT ) was used. The first trial for image reconstruction was done in a workstation (Windows 10) consisting of Xeon, 256 Gbyte RAM, NVIDIA TESLA P100 32GB, but the MATLAB (R2017b) was suddenly shutdown. It looked kind of memory issue.

Did anybody have similar experiences? and could anybody identify which one is the issue, RAM or GPU? If possible, advise minimum requirements of the GPU machine for this task, please.

Thanks!

[Voss]

A matrix of that size and class (single complex) will take up about 127 GB of RAM, which is pretty close to half your total RAM, so if you have any operations in your code where MATLAB is going to make a temporary copy of that matrix, you're very likely to run out of memory. (The extended matrix size would take about 9000 GB, in which case you're guaranteed to run out of memory.) My guess would be that it was indeed a memory issue, and it might be a good idea to figure out an alternative method whereby the entire matrix doesn't need to be loaded into RAM at one time (and such a method will be required if you move to the extended matrix size).

However, MATLAB doesn't usually shutdown suddenly when it runs out of memory; it usually throws an Out Of Memory error, so I could be wrong. I'm mostly speculating based on the specified matrix size and amount of RAM.

[Kyungpyo Hong]

Thanks for your answer.

The same matrix size (384x384x384x30x5x2) was tested in another computer consisting of Intel i9, 64 Gbyte RAM and GeForce GTX 1080, and MATLAB threw out of memory: Total available memory: 8589934592, Required memory: 8652791808.

Is this GPU memory or RAM? Looks like GPU to me, because GeForce GTX 1080 has 8 GB memory. Could you identify which one is? If it is GPU, I am wondering why the same matrix was unable to implement in better environment such as TESLA P100 32GB.