当前位置:网站首页>Nonvolatile MRAM memory used in all levels of cache
Nonvolatile MRAM memory used in all levels of cache
2020-11-06 22:00:00 【DISPLAY】
Magnetoresistive random access memory MRAM It's a new type of memory , Its advantages include fast reading speed, high integration and non-volatile . At present, a lot of research is mainly devoted to the research of MRAM Used in computer storage system . At the same time, nonvolatile MRAM Memory is also used at all levels of cache .
MRAM replace SRAM do L2 Cache
First compare the same area of MRAM and SRAM. Directly use the same area of MRAM Replace SRAM do L2 Caching can reduce the error rate . But the write latency is long . When the write operation strength is high , The advantage of reduced error rate will be offset by long delay, resulting in performance degradation . Although this direct substitution can greatly reduce the leakage power consumption , But when write intensive , Significant increase in dynamic power consumption , Make the effect of reducing energy consumption worse . If you directly use the same area of MRAM replace sram , When write operations are intensive , Its disadvantages such as long write delay and high energy consumption will offset its advantages .
MRAM As L3 Cache
L2 Too much capacity increases access latency, so it is not applicable . Add a level to the storage system L3 The feasibility of caching . The researchers calculated a 128MB,4-banks,16-way,256-byte block The cache area is only 161mm2, Suitable for stacking on current processors . The time model shows that the delay is only 15.82ns, Much less than the average memory access time . In different situations IPC The speed has increased 0.03% To 108%. Yes L2 The high cache error rate is greatly improved . And this improvement only needs to be added 0.4W Power consumption of .
MRAM Used as main memory
Research has shown that stacking on a chip DRAM The feasibility of memory . And DRAM Compared with MRAM There is no need to refresh periodically . But it's still DRAM The highest degree of integration . Current stack DRAM To improve the performance of 19%( For integers ) and 40%( For floating-point numbers ). We have reason to believe that stacking MRAM Technology will perform better because of its shorter latency .
although MRAM The low power consumption makes it possible to realize multi-layer stacking without worrying about the temperature , But the delay will increase , And too many stacking layers will lead to a decrease in yield . So the current MRAM The technology is not mature enough to be applied to main memory ( Because it's not big enough ) , But it can be used in embedded devices with special requirements for low power consumption .
版权声明
本文为[DISPLAY]所创,转载请带上原文链接,感谢
边栏推荐
- 【涂鸦物联网足迹】物联网基础介绍篇
- An article taught you to download cool dog music using Python web crawler
- Exclusive interview of guests at | 2020 PostgreSQL Asia Conference: Wang Tao
- September 9, 2020: naked writing algorithm: two threads print numbers 1-100 in turn.
- What kind of music do you need to make for a complete game?
- 1万辆!理想汽车召回全部缺陷车:已发生事故97起,亏损将扩大
- 2020-09-03:裸写算法:回形矩阵遍历。
- Python 100 cases
- Unexpected element.. required element
- Windows 10 蓝牙管理页面'添加蓝牙或其他设备'选项点击无响应的解决方案
猜你喜欢

Big data processing black Technology: revealing the parallel computing technology of Pb level data warehouse gaussdb (DWS)

2020-08-14:数据任务的执行引擎用的哪些?

Common mathematical basic formulas of recursive and backtracking algorithms

应用层软件开发教父教你如何重构,资深程序员必备专业技能

To solve the problem that the data interface is not updated after WPF binding set

All the way, I was forced to talk about C code debugging skills and remote debugging

What are the highlights of Huawei mate 40 series with HMS?

image operating system windows cannot be used on this platform

An article will take you to understand CSS alignment

Unity performance optimization
随机推荐
2020-08-15:什么情况下数据任务需要优化?
MRAM高速缓存的组成
[elastic search engine]
JVM memory allocation - xms128m - xmx512m - XX: permsize = 128M - XX: maxpermsize = 512M
Summary of front-end performance optimization that every front-end engineer should understand:
2020-09-09:裸写算法:两个线程轮流打印数字1-100。
What is the tensor in tensorflow?
消防器材RFID固定资产管理系统
August 24, 2020: what are small documents? What's wrong with a lot of small files? How to solve many small files? (big data)
An article will take you to understand SVG gradient knowledge
C calls SendMessage to refresh the taskbar icon (the icon does not disappear at the end of forcing)
What kind of music do you need to make for a complete game?
File download manager realized by electron
2020-08-20: the difference between go and python?
Why is quicksort so fast?
August 14, 2020: what are the execution engines for data tasks?
Windows 10 蓝牙管理页面'添加蓝牙或其他设备'选项点击无响应的解决方案
Countdown | 2020 PostgreSQL Asia Conference - agenda arrangement of Chinese sub Forum
What the hell is fastthreadlocal? The existence of ThreadLocal!!
1万辆!理想汽车召回全部缺陷车:已发生事故97起,亏损将扩大