SPEC CPU2017 is an industry standard benchmark suite. OEMs use it to set performance expectations for their systems, and CPU make often use it to tune

Running SPEC CPU2017 on Chinese CPUs, and More

submited by

Style Pass

2024-10-19 19:00:07

SPEC CPU2017 is an industry standard benchmark suite. OEMs use it to set performance expectations for their systems, and CPU make often use it to tune their designs. We've published some SPEC CPU2017 estimated results, and now it'll be interesting to go back and run SPEC CPU2017 on CPUs we've gone over in prior articles. Chinese CPUs are especially interesting because they're often not tested by more mainstream tech outlets, so it's a good place to start. Some other CPUs were tested to provide comparison data.

Results here are estimated because SPEC CPU2017 has a long list of requirements. We've set out to meet all the technical requirements like completing all tests in a suite from one runcpu invocation and using a single file system. Differences primarily come down to documentation requirements. As before, we're using GCC 14.2.0 and running bare metal Linux. GCC 14.2.0 is either compiled from source or run through a Debian chroot if it's not available from the distribution's packages. Optimization flags are set to let the compiler perform a typical level of optimization while targeting the tested CPU's ISA extension. For example, that's -O3 -march=native -mtune=native -fomit-frame-pointer for x86-64, or -O3 -mcpu=native -fomit-frame-pointer for aarch64.

We'll also focus on single threaded performance with SPEC CPU2017 rate tests running a single copy. There's a small exception for evaluating SMT gains, which will be tested with two copies pinned to sibling threads on a single core.