AVX-512 Performance With 256-bit vs. 512-bit Data Path For AMD EPYC 9005 CPUs
Right from the start we see very clearly the added performance that AMD EPYC 9005 series can provide from a 512-bit FP data path compared to the 256-bit "double pumped" approach used with prior EPYC 9004 series processors. EPYC Zen 4 was nice for having AVX-512 on a 256-bit data path and proved to be efficient but now with the Zen 5 server processors is that much more performance potential from a native 512-bit data path.
In the case of the miniBUDE benchmark, using AVX-512 was more efficient than when artificially disabling it for the CPU. As for the power impact from 256 vs. 512 bit data path, there was very little difference and the performance gains of the FP512 path proved to deliver the best performance-per-Watt.
Unlike early AVX-512 implementations on Intel Xeon processors, there was no negative CPU temperature impact from using AVX-512.
NAMD enjoyed a nice 10% bump in performance from the 512-bit data path on this EPYC 9754 processor. Had AMD not picked up AVX-512 support in Zen 4, they would have not been nearly as attractive for HPC workloads as they are now.
AVX-512 with the native 512-bit data path route proved to deliver the best performance-per-Watt.
There's a clear difference in AVX-512 performance as well with the wonderful GROMACS software.