Matrix multiplication in PyCuda - what is the reason for the increase in error?

Question

Python

Having tested this example on pycuda, there is an increase in the error with an increase in the size of the matrix (which is theoretically logical, since more floats are added).
those. np.allclose(c_cpu, c_gpu.get()) returns false.
although I did not understand this limitation

40 # define the (square) matrix size
41 # note that we'll only use *one* block of threads here
42 # as a consequence this number (squared) can't exceed max_threads,
43 # see documen.tician.de/ pycuda/util.html#pycuda.tools.De...
44 # for more information on how to get this number for your device
45 MATRIX_SIZE = 2

related: stackoverflow.com/questions/4104010/cuda-float-poi...

Answer 1

Try to run this test for half an hour:
wili.cc/blog/entries/gpu-burn/gpu_burn-0.4.tar.gz
In my experience, gaming video cards after a couple of years of use may not pass this test, in other words, continuing to work in toys, may be miscalculated.