alt Hacker NewsThe SIMD stuff is incredible. I have been having lots of fun with it. You can use LLMs as a scalar to SIMD transpiler, it works amazingly well.
Sure a SIMD expert writing assembly can probably do a better job than an LLM using these new intrinsics, but it’s still massively faster.
Replies
nkanaev • yesterday at 8:58 PM
Agreed. I've recently translated a pangram generator project written in Rust [1] leveraging SIMD to do the same in Go [2] to see how it fairs in terms of speed - the results are pretty close. In my local machine I'm getting ~3GHz in Rust vs ~2.4GHz in Go, which I think is really impressive.
drivebyhooting • today at 2:08 AM
“Do a breakthrough and make this SIMD fastest”
I still believe that SIMD support is one of the most underrated new features in Go.
It's relatively straightforward to read and to write code using Go's SIMD package (the caveat being that you have to convert your data to SoA manually), and it gives comparable performance to other languages, since there's little in the way of GC overhead, bounds checking, etc, in this case.
So SIMD not only increases performance on its own, but it also closes the performance gap between Go and C++ / Rust, which has been a major cause for rewrites in the past. The memory usage overhead due to GC doesn't go anywhere of course, so there are still performance reasons to "Rewrite in Rust", but it's now become much easier to just optimise the hell out of Go code instead.