3 ms·
Well, AMD currently has better CPUs for high-performance computing, so OP's users would benefit from this public service announcement, if anything.
by gnulinux 8y ago
Well, AMD currently has better CPUs for high-performance computing, so OP's users would benefit from this public service announcement, if anything.
- Symmetry 8y agoI thought it was the opposite of that? That is, AMD chips are currently great for most workloads but Intel still has a definite edge in working with big vectors of floating point numbers of the sort you usually encounter in HPC. Mostly because their core's vector units are 512 instead of 256 bits wide.
- celrod 8y agoSkylake-X and certain Xeons have avx-512 (which includes two 512 bit fma units). The rest only have 256 bit wide vectors, like Ryzen. But they still have the advantage of two 256 bit fma units, while Ryzen instead relies on two 128 bit fma units, meaning the fma instructions critical to matrix operations are faster on Intel. I think the idea for HPC though is that you want to offload these highly vectorizable operations to a GPU. Or maybe you're doing a lot of Monte Carlo that it is hard to vectorize. I really do like avx-512 though. If you're writing a lot of your own code, and that code involves many small-scale operations and has control flow (like in many Monte Carlo simulations), it's a lot easier than mucking with a GPU. If you're using gcc though, be sure to add `-mprefer-vector-width=512`, otherwise it will default to 256 bit vectors (clang and Julia use 512 by default).