Error using gpuArray Maximum variable size allowed on the device is exceeded.

>> X = gpuArray(X);

>> Error using gpuArray

>> Maximum variable size allowed on the device is exceeded.

My X has the size

[230227128 10] = size(X)

230227128 x 10 > 2147483647 == int32

I am using

R2019a (9.6.0.1072779)

64-bit (win64)

March 8, 2019

and GPU has 45 GB (RTX 8000 48 GiB).

I am aware that there is a hard limit of GPU memory usage in Matlab.

But it is indeed discouraging and I am trying to write a C++ code to circumvent it.

Can there be a hotfix on this?

Answer 1

[Joss Knight]

The NVIDIA libraries that MATLAB uses to do GPU computation store array lengths as 32-bit integers and therefore MATLAB cannot use any of those libraries for arrays with more than intmax('int32') elements. Therefore this is a hard limit on the array size. Try dividing your computation up.

[HyoSeung Kang]

How stupid I am since it took very long time to truly understand that I can really split the data into smaller pieces. Thanks for the advice.

[Subhash Chandra Ranga]

May I know how to split the data then feed it to a function and again combine the result

[John Lee]

I am experiencing the same problem. I just started to use a A100 80GB PCIe anticipating to calculate eig() on a single very large matrix close to half of the GPU memory size. And I found this post.

The gpuArray size limit of intmax('uint32') seems to be too restrictive these days. It means I could just use 3090Ti or even 1080Ti since I don't use many gpuArray variables simultaneously.

Is there a MATLAB's plan to increase the gpuArray size limit to intmax('uint64'), or

is there a multi-step numerical algorithm to make the input matrix to eig() smaller and achieve the same end result (I doubt)?