HBM: High-Bandwidth Mistake

The transformer architecture is winning based on brute force alone so why would the industry pick an optimal but exotic solution on the hardware side?

It sounds to me like brute force is going to stay on the software side and with that hardware side brute forcing is a foregone conclusion.

If optics become a thing, the first GPU memory implementations would just switch out the signal bumps with vertically integrated optical channels.

The author isn't appreciating how ancient of a technology HBM is.