CARM Tool brings universal CPU benchmarking to Intel, AMD, ARM, RISC-V

As HPC systems and CPU architectures grow increasingly complex, developers struggle to identify performance bottlenecks. The Cache-Aware Roofline Model (CARM) offers intuitive guidance by visualizing computational vs. memory limits, but until now, automatic CARM tools were vendor-specific (Intel Advisor) or underdeveloped (ARM), with no support for AMD or RISC-V. A new paper from Universidade de Lisboa introduces CARM Tool, an open-source automatic benchmarking and application analysis framework that fills this gap. It delivers universal CARM support across all major ISAs—x86 (Intel and AMD), ARM, and RISC-V—using custom assembly microbenchmarks tailored to cover the full performance spectrum of modern CPUs, from scalar operations to all supported vector ISA extensions, and every level of the memory hierarchy.

CARM Tool integrates application analysis via hardware performance counters and dynamic binary instrumentation, allowing developers to pinpoint exactly where their code hits architectural limits. Experimental results show that the constructed roofline roofs deviate less than 1% from theoretical maximums across tested architectures, validating the tool's accuracy. By automating the otherwise tedious process of building roofline models, CARM Tool empowers engineers to optimize code for any CPU—without relying on proprietary tools. The research was presented at the 2024 International Symposium on Workload Characterization (IISWC '24) and is now available on arXiv.