| Texto completo | |
| Autor(es): |
Número total de Autores: 4
|
| Afiliação do(s) autor(es): | [1] Univ Estadual Campinas, Inst Comp, Campinas - Brazil
[2] Fed Univ Technol, Parana Campus, Campo Mourao - Brazil
Número total de Afiliações: 2
|
| Tipo de documento: | Artigo Científico |
| Fonte: | FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE; v. 113, p. 556-570, DEC 2020. |
| Citações Web of Science: | 0 |
| Resumo | |
Approximate memories expose data elements to errors in order to improve energy efficiency. For a large fraction of data, these errors are inconsequential or lead only to small losses in application output quality. Nevertheless, for some critical data, errors may lead to execution flow crashes, resulting in non-produced outputs and wasted computational and energy resources. Thus, these techniques require some level of control over approximations to generate acceptable outputs and, consequently, to maximize the energy benefits. Many proposed interfaces for approximate memories rely on burdensome instrumentation of the program or on user annotations to protect critical data. We present AxRAM, a lightweight interface for approximate data that avoids crashes without user annotations. AxRAM relies on a memory with configurable reliability levels and protects from errors critical data regions commonly found on a number of applications. Furthermore, our interface implements a resilient addressing scheme that reduces invalid data accesses that lead to execution crashes. In an embedded computing scenario with a dual-VDD SRAM, our implementation of AxRAM reduces 51% of the execution crashes compared to an unprotected approximate memory, resulting in energy savings for 9 out of 12 profiled applications. (C) 2020 Elsevier B.V. All rights reserved. (AU) | |
| Processo FAPESP: | 18/24177-0 - Suporte Arquitetural para Computação Aproximada |
| Beneficiário: | Isaías Bittencourt Felzmann |
| Modalidade de apoio: | Bolsas no Brasil - Doutorado |