top of page
MATMUL
This visual riff on a matrix multiplication is based on Simon Bohem's description meant to help “understand the most important performance characteristics of the GPUs that are used for modern deep learning. This includes coalescing global memory accesses, shared memory caching and occupancy optimizations”.
Specifically, this matrix of cubes is meant to represent the blocks and threads that make up the grid of kernals in a CUDA matmul code. As one looks deeper into it, there are categories of same colours, and colour combinations. This lends itself to be a bit of a visual riddle, to find like patterns/colours.
Similarities to the bounding boxes around faces and items in AI vision are unintentional, but impossible to ignore once considered. These were a 'happy accident' that was revealed only after completion of the piece
bottom of page