I always look forward to reading a post by Lemire, they're usually about getting the best performance out of modern processors using SIMD.
This is a test of AVX-512 code, using the simdutf library. The AMD processor manages 11GB/s against Intel's 6GB/s. The AMD chip also retires 4 Instructions per cycle vs intel's 2.6.
About the simdutf library:
> It is used by major browsers and JavaScript runtimes like Node.js or Bun. A common operation that matters is the conversion from UTF-16 to UTF-8. Internally, JavaScript relies on UTF-16, thus most characters use 2 bytes, whereas the Internet defaults on UTF-8 where characters can use between 1 and 4 bytes.
TinkersW•2mo ago
I am surprised Intels server chips can only do 2 AVX512 ops per cycle, that is rather sad given how long they have supported it in server chips, and I hope isn't a sign of things to come with Nova Lake.
kristianp•2mo ago
This is a test of AVX-512 code, using the simdutf library. The AMD processor manages 11GB/s against Intel's 6GB/s. The AMD chip also retires 4 Instructions per cycle vs intel's 2.6.
About the simdutf library:
> It is used by major browsers and JavaScript runtimes like Node.js or Bun. A common operation that matters is the conversion from UTF-16 to UTF-8. Internally, JavaScript relies on UTF-16, thus most characters use 2 bytes, whereas the Internet defaults on UTF-8 where characters can use between 1 and 4 bytes.