A 28-nm 8-bit Floating-Point Tensor Core-Based Programmable CNN Training Processor With Dynamic Structured Sparsity | IEEE Journals & Magazine | IEEE Xplore