deeplearning Matlab is not compatible with graphics card

i use parallel.gpu.enableCUDAForwardCompatiblity function. but i found when network input size is 1*400*1,train samples number is 16800.batchsize even 1000,i can train,and get a good result.but the other network,input is 1*60*1,when batchsize maybe over 256,than network be very wrose.acc is 1/(number of class).ie i have 4 clsass,then is 25%.batchsize lower 256,nertwork works very well. when input is 50*64*1,then any batchsize is wrose.

my matlab is 2020b.graphics card is 3090.

[Rik]

It is not clear to me what exactly is you question. Have a read here and here. It will greatly improve your chances of getting an answer.

Also, given the current availability (or rather the lack thereof) of the RTX 3090 I doubt many people will be able to replicate specific hardware-related issues.

[bob james]

the question maybe :the same network,same train data,the same model parameters,i use cpu can get very good results.buy when using gpu, the results was very wrose.

Answer 1

[Joss Knight]

As explained in the documentation, both for CUDA support and for the enableForwardCompatibility function, "Enabling forward compatibility can result in wrong answers and unexpected behavior during GPU computations".