Microarchitectural comparison and in-core modeling of state-of-the-art CPUs

The paper titled "Microarchitectural comparison and in-core modeling of state-of-the-art CPUs: Grace, Sapphire Rapids, and Genoa" by Jan Laukemann, Georg Hager, and Gerhard Wellein examines the performance of leading CPUs from Nvidia, AMD, and Intel in the high-performance computing (HPC) sector. The authors develop an in-core performance model for the microarchitectures Zen 4, Golden Cove, and Neoverse V2, utilizing the Open Source Architecture Code Analyzer (OSACA) tool and comparing it with LLVM-MCA. The study highlights the unique features and performance characteristics of each CPU, particularly focusing on the "write-allocate (WA) evasion" feature, which minimizes memory traffic from write misses. The findings indicate that the Grace Superchip has an optimal implementation of WA evasion, while Zen 4 requires explicit non-temporal stores to avoid write allocates. The research includes various microbenchmarks and evaluates the capabilities of full nodes, providing insights into the competitive landscape of modern CPUs.

- The study compares the performance of Nvidia's Grace Superchip, AMD's Genoa, and Intel's Sapphire Rapids CPUs.

- An in-core performance model is created for the microarchitectures Zen 4, Golden Cove, and Neoverse V2.

- The "write-allocate evasion" feature is a key focus, showing Grace's superior implementation.

- The research utilizes the Open Source Architecture Code Analyzer (OSACA) and LLVM-MCA for performance analysis.

- Findings suggest that Zen 4 requires specific programming techniques to optimize memory traffic.