Hacker News new | past | comments | ask | show | jobs | submit login

I don't know the details but the WebGL backend will have severe limitations because it doesn't use compute shaders or CUDA. This means that certain functionality like random writes to arbitrary buffers can only be emulated through workarounds that are like an order of magnitude slower than using compute shaders, and some things are going to be impossible entirely. There is something about CUDA in the link you provided but since that isn't natively supported by browser either, it will require the user to install something or use a server backend where communication between server and client is going to be an excruciatingly slow bottleneck.

So no, this isn't a web thing and not usefull for web apps/web pages, except maybe for a very limited and small set of use cases.

When I last benchmarked tensorflow.js, it was 40 times slower than native tensorflow on my laptop. However, even if WebGL compute shaders were available, Nvidia's cuBLAS and cuDNN libraries would still be faster.

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