Linux Patch Targets Full-Width AVX-512 for Faster ChaCha on Zen 5
Previously, the relevant AVX-512 code used 256-bit registers to process up to 8 data blocks at a time, avoiding severe downclocking on early Intel Skylake systems running 512-bit instructions. [[fact:fact-9fa673zlz]] StrongSwan developer Martin Willi added a full-width AVX-512 function to libcrypto's ChaCha code that processes 16 data blocks. [[fact:fact-br7euwmyn]] Linux's ChaCha implementation has long supported AVX-512 but remained constrained, and the issue is expected to be addressed in Linux 7.5. [[fact:fact-66la9bd99]] The patch is queued in the libcrypto-next branch and is expected to enter the mainline during the Linux 7.5 merge window. [[fact:fact-drblx8sh2]] Estimates indicate that ChaCha performance could improve by up to 99% on AMD Zen 5 platforms with a full 512-bit data path. [[fact:fact-ng1qiezzk]]