Hacker News Reader

Comment by bsaul | Hacker News Reader

A bit off topic, but does anyone knows if AI ( aka modern neural networks) plays a role in cpu design nowadays ?

not sure if they play a role in designing CPUs, but AMD says that their Zen CPUs have a neural network inside for branch prediction [1] which is not something new [2] [3]

[1] https://www.amd.com/en/technologies/sense-mi

[2] https://www.jilp.org/cbp2014/paper/DanielJimenez.pdf

[3] http://cseweb.ucsd.edu/~atsmith/rnn_branch.pdf

CPU layout and routing optimizations are global discrete optimization problems. Neural networks don't have much edge on these problems. Simulated annealing is the most used optimization method in routing and layout tools.

There are number of research papers that attempt to use neural networks for combination optimization, but no interesting results.

Neural networks aside, machine learning is often used for implementing prediction mechanisms in CPUs. Good examples of prediction mechanisms are:

- Branch Predictor

- Cache Replacement Policy

- Memory Prefetcher

- Schedulers

In the designing of CPUs or if they are designed for running various neural networks? Because Nvidia seemed to be building giant products aimed pretty much exclusively at that. Google has their own "TPUs" for that. Intel's next Phi is designed for running that stuff.

As for the design of CPUs, they do some automated layouts and some things like that, but I think that might not be what you mean.

Yes, i mean in the designing of CPUs. Like transistor layout. It seems like a big problem space where optimal arrangement is hard to find. So i wondered if people trained NN to do that.

Probably not. Do you use AI to write code?

A compiler does a lot of AIish things in turning my C code into a sequence of instructions. Much more classical AI than machine learning but still within the AI umbrella.

I was referring to the RTL-based IP that goes into each subsystem of the CPU, hence the analogy to code. You're talking about compiler instruction scheduling, a lot of which are a bunch of non-AI algorithms. If I'm missing something, I would appreciate the references to the functionality you're referring to.

Most of the time not even the natural one

Not now, but i'm pretty sure this would be a good technic to try for next gen autocomplete.

Here's one CPU that has nn in mind:

https://cloudplatform.googleblog.com/2017/04/quantifying-the...

Reply on news.ycombinator.com