Samsung's New HBM2 Memory Thinks for Itself: 1.2 Tflops of Embeded Compute Power

This actually seems quite useful and innovative.

Most of the processing these days involves moving buffers: networking, databases, graphics, and core parts of most OSes would involve a significant amount of slapping buffers together, changing them for which a CPU might not be that efficient for IO bound operations even with DMA or other fancy instructions.

AI/ML buzzword notwithstanding, this is very cool.