首页|期刊导航|计算机工程与科学|基于延迟写回缓存的多粒度片上访存容错设计

基于延迟写回缓存的多粒度片上访存容错设计OA

A multi-grained on-chip memory access error-tolerant design based on delayed write-back cache

中文摘要英文摘要

片上存储器是现代微处理器架构中至关重要的组成部分,其可靠性和性能直接影响着整个系统的稳定运行和运算效率.随着片上存储器集成度和性能需求的不断增长,数据访存设计面临着可靠性、吞吐量及高能效的巨大压力.传统的粗粒度访存模式虽可降低冗余容错存储和硬件开销,但在空间局部性较差的应用场景中,其可靠性和访存带宽性能难以发挥.为了支持多种数据粒度访存的可靠性加固设计,提出了一种基于延迟写回缓存(DWB-cache)的高能效细粒度访存容错处理设计,保留了粗粒度访存的低冗余存储器面积开销的容错优势,同时采用 DWB-cache 隐藏了细粒度数据访存的处理延迟,有效兼顾了硬件开销和访存性能.将该设计应用于某高性能微处理器的标量存储器 M64-SM 中,实现了细粒度延迟写回纠错模块(FGDWB-ECC),解决了访存流水线的数据一致性问题,探索了最优 DWB-cache 深度.实验结果表明,相较于最细粒度访存加固的设计基准,采用 FGDWB-ECC 模块使得 M64-SM 整体面积开销及其冗余存储器面积开销分别降低了 15.4%和 80.9%;在 GEMM 和 FFT 典型应用场景中,该模块设计的性能表现与基准设计持平.

On-chip memory is a critical component in modern microprocessor architectures,and its reliability and performance directly impact the stability and computational efficiency of the entire sys-tem.With the continuous increase in the integration density and memory access performance,on-chip memory faces significant challenges in achieving high reliability,high throughput,and energy-efficient design.While traditional coarse-grained access modes reduce error-tolerant redundancy and hardware overhead,their reliability and memory bandwidth perform poorly for applications with weak spatial locality.To address this,we propose a high-efficiency,fine-grained delayed write-back memory access error-tolerant design based on delayed write-back cache(DWB-cache).This design combines the low error-tolerance redundancy benefits of coarse-grained access with the performance benefits of fine-grained access,using DWB-cache to hide fine-grained data processing latency and balancing hardware overhead with access performance.We implemented FGDWB-ECC module in the M64-SM scalar memo-ry of a high-performance microprocessor,solving the DWB-cache consistency issue and determining the optimal cache depth.Experimental results demonstrate that,compared to the finest-grained error-tolerance baseline,FGDWB-ECC module achieves a 15.4%reduction in total area overhead and an 80.9%reduc-tion in redundant memory overhead.Moreover,application performance in both GEMM(good spatial locality)and FFT(poor spatial locality)remains on par with the baseline.

郭岚婷;陈海燕;李晨;刘胜

飞腾信息技术有限公司,湖南长沙 410073国防科技大学计算机学院,湖南长沙 410073||先进微处理器芯片与系统教育部重点实验室,湖南 长沙 410073国防科技大学计算机学院,湖南长沙 410073||先进微处理器芯片与系统教育部重点实验室,湖南 长沙 410073国防科技大学计算机学院,湖南长沙 410073||先进微处理器芯片与系统教育部重点实验室,湖南 长沙 410073

信息技术与安全科学

片上存储器访存容错多粒度访存延迟写回缓存细粒度写SRAM

on-chip memoryaccess error-tolerancemulti-grained memory accessdelayed write-back cachefine-grained write accessstatic random-access memory(SRAM)

《计算机工程与科学》 2026 (9)

1548-1559,12

10.3969/j.issn.1007-130X.2026.09.003

评论