Pyxis

专用加速器在特定应用领域提供性能和效率增益。稀疏数据结构或/和表示存在于广泛的应用程序中。然而，为稀疏应用程序设计加速器具有挑战性，因为没有架构或性能级别的分析模型能够完全捕获稀疏数据的频谱。加速器研究人员依靠实际执行来为他们的设计获得精确的反馈。在这项工作中，我们提出了 PYXIS，这是一个针对稀疏数据的专用加速器的性能数据集。 PYXIS 收集加速器设计和实际执行性能统计数据。目前，PYXIS 中有 73800 个实例。 PYXIS 是开源的，我们通过新的加速器设计和性能统计数据不断发展 PYXIS。 PYXIS 可以使加速器、架构、性能、算法和许多相关主题领域的研究人员受益。Dedicated accelerators deliver performance and efficiency gains in specific application domains. Sparse data structures and/or representations exist across a wide range of applications. However, designing accelerators for sparse applications is challenging, as no analytical models at either the architectural or performance level can fully capture the spectrum of sparse data. Accelerator researchers rely on real-world execution to obtain precise feedback for their designs. In this work, we present PYXIS, a performance dataset for dedicated accelerators targeting sparse data. PYXIS collects accelerator design parameters and real-world execution performance statistics. Currently, there are 73,800 instances in PYXIS. PYXIS is open-source, and we continuously evolve it with new accelerator designs and performance statistics. PYXIS can benefit researchers across the fields of accelerators, computer architecture, performance analysis, algorithms, and numerous other related domains.

搜集汇总

背景与挑战

背景概述

Pyxis是一个针对稀疏数据专用加速器的性能数据集，包含73800个实例，旨在通过收集实际执行性能统计数据来解决稀疏应用加速器设计缺乏分析模型的问题。该数据集由加州大学于2021年开源发布，可支持加速器、架构及相关领域的研究，并持续更新发展。