AI ASIC的基准测试、优化和生态系统协作的整合 | KC23

Consolidation in Benchmarking, Optimization, and Ecosystem Collaboration for AI ASICs - Minghui Yu

AI ASIC的基准测试、优化和生态系统协作的整合 | Consolidation in Benchmarking, Optimization, and Ecosystem Collaboration for AI ASICs - Minghui Yu, ByteDance

ASIC在AI加速中越来越受欢迎。然而，对于IT公司来说，采用新的ASIC并不容易。采用新的ASIC是一项耗时的工作，需要跨团队沟通、模型选择以及交付预期的性能和准确性。不透明的编译要求以及对给定ASIC的模型熟练度或不熟练度造成了巨大的负担。为了解决这些问题并加快评估过程，ByteMLPerf从生产角度开发出来，以协助评估ASIC。它专注于软件和硬件的易用性和多功能性，通过调整其基准测试工具来使模型和运行时在实际应用场景中保持一致，并将编译能力作为一流的API内置，从而显著提高可重复性。该工具提供了广泛的指标，以更好地反映实际情况进行综合评估。此外，编译后端最大限度地发挥了每个模型的性能潜力，并更好地利用了ASIC。

ASICs are becoming more and more popular in AI acceleration. However, adopting new ASICs is not trivial for IT companies. It’s time-consuming to adopt new ASICs, requiring cross-team communication, model selection, and delivery of expected performance and accuracy. Huge burdens are caused by opaque compilation requirements, and model proficiency or incompetence for a given ASIC. To address these issues and expedite evaluation, ByteMLPerf is developed to assist in assessing ASICs from a production perspective. It focuses on the ease of use and versatility in software and hardware, tailoring its benchmarking tool to align models and runtime in real-world use cases, and building-in compilation capability as a first-class API, significantly improving reproductivity. With a wide range of metrics, the tool provides a comprehensive evaluation to better reflect real-world conditions. Also, the compilation backend maximizes the performance potential of each model and leverages ASICs better.

CNCF概况（幻灯片）

扫描二维码联系我们！

CNCF (Cloud Native Computing Foundation)成立于2015年12月，隶属于Linux  Foundation，是非营利性组织。 

CNCF（云原生计算基金会）致力于培育和维护一个厂商中立的开源生态系统，来推广云原生技术。我们通过将最前沿的模式民主化，让这些创新为大众所用。请关注CNCF微信公众号。