Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I suspect that there will be a LOT of TPU e-waste in about 3-5 years. One can call me reactionary, but I see a greater benefit in providing SIMD primitives for known bottlenecks, co-processing for known inner loops, than to continually jump through the hoop of committing passing trends to hardware.

I suspect thst Microsoft will implement brilliantly on a bad idea, but that it still won't gain much traction, as Co-pilot is not the code assistant we need.

But since hammers are in fashion, expect attempts to shoehorn the world into nails.




If vendors will go for strictly on-die TPUs, then maybe. However, I expect these machines to be quite capable without the AI hype.

I hope ar least some laptop vendors will use PCIe cards for their TPUs so that there's an upgrade path of sorts, and reusing older cards for simpler models remains possible.


LLMs aside I would guess we may see more modest use cases for TPU/NPUs on PC. Phones have been coming with them for several years now - granted with stronger use cases around AR, photo retouching, and voice assistants. Maybe you are right though.




Consider applying for YC's Fall 2025 batch! Applications are open till Aug 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: