Note that the following benchmarks are for the standard FIPS-203 API.
In particular, the public and private keys are stored in packed form and are subject to input validation.
In applications where the private key can be retained in expanded form (e.g., TLS), performance of mlkem-native will be much better.
For benchmarks on x86 TurboBoost and SMT are enabled (as we are running on shared AWS instances). SMT may lead to slight performance variations (<3%) over mutliple runs. TurboBoost results in lower cycle counts. For comparison to other implementations, we recommend turning both SMT and TurboBoost off.