AbstractBenchmark suites have become the most important wheel to drive the research of GPU architecture designs. The most popular GPU benchmark suites, such as Rodinia and Parboil, cover a wide range of real-world applications, which provide a good reference for the researchers to estimate the performance behaviors of the GPU devices in production. However, the existing GPU benchmarks were initially proposed a decade ago, which unfortunately cannot reflect the unique features of the latest GPU applications accurately. Therefore, it is inappropriate to employ the stale benchmarks as the reference to evaluate the state-of-the-art GPU architectures.
Tackling this challenge, we propose BcBench, a new set of GPU workloads collected from the emerging blockchain applications. BcBench is designed for estimating the performance behaviors of the future GPU architectures in accelerating the execution of the emerging applications. To this end, we have ported BcBench to the popular GPU simulators (e.g., GPGPU-Sim) for in-depth performance analysis. We first characterize the blockchain workloads at the micro-architecture level of GPUs fully by leveraging the adequate statistics generated by the GPU simulator. We then conclude five key observations from the workload characterization. We further explore five future GPU architecture designs, which target for the blockchain applications.