PCEVA,PC绝对领域,探寻真正的电脑知识
开启左侧

转:内存各时序对性能的影响(多图)

  [复制链接]
royalk 发表于 2010-12-11 18:05 | 显示全部楼层 |阅读模式
点击数:16293|回复数:31
之前写过一些文章提到内存超频只是辅助CPU超频,单独超内存本身并不会得到太多性能提升,但是却没有拿出实际的数据比较,刚好看到国外的一篇关于这方面的测试,转过来给大家看看:
原文:http://www.xtremesystems.org/forums/showthread.php?t=263183
这人英语表达能力不太好,大家不用看文字说明了,看图就能说明一切

测试平台
AMD Phenom II X6 1090T
ASUS M4A89GTD PRO/USB3
G.SKILL PIS PC3-19200 CL9 4GB (2GB x 2) @ 2000MHz
Intel X25-M G2 80GB
PC Power & Cooling Silencer 910W 80PLUS Silver
测试设定:


测试项目:
1. CL: 6~11
2. tRCD: 9~11
3. tRP: 6~11
4. tRAS: 24~30
5. Command Rate: 1 / 2
6. tRC: 15~40
7. tRFC: 90ns / 110ns / 160ns / 300ns 350ns
8. tREF: 3.9ms / 7.8ms

测试软件项目:
SuperPi 32M
Everest内存带宽、延迟
H264编码(不使用GPU编码)

Super PI test

Note: sorry for 'magic graph'. since it takes numerous min. for 32M Pi caculation, results are differred only by a very second. so it's unavoidable to show 'meaningful' graphs.

1. CL Test


2. tRCD Test


3. tRP Test


4. tRAS Test


5. Command Rate Test


6. tRC Test


7. tRFC Test


tRFC and tREF are indexed by actual time (second) while others by clock count. It's because they directly indicate characteristics of electrical working mechanism of a memory chip. Both are related to a memory's refresh rate, tREF determines how frequently a chip shall be refreshed and tRFC determines the time interval from the moment of 'refresh' (memory access is unavailable) to when a memory access is back to available. So in case of tREF, the bigger value the better at performance, but the contrary is more likely at stability.

8. tREF Test


As of now, we covered all 8 elements. Let's see overall analysis.

Overall Analysis: Super PI


(...sorry for Korean letter. actually I uploaded this bench to my private blog) The title of the graph is that 'how a ram timing value affects performance when we change just '1-degree' of ram timing option provided by BIOS'. As you see, tREF affects the most but not more than 1%, and followed by tRFC-Command Rate-CL. It's impressive that well-known items such as CL, tRCD affect far less than 'minor' items such as tREF and tRFC.

Then let's go to next test result - Everest Cache & Memory Benchmark


Everest Cache & Memory Benchmark
- Memory Bandwidth test -

Each Everest test consists of two parts: Bandwidth and Latency.
Let's cover the bandwidth results first.

1. CL Test


2. tRCD Test


3. tRP Test


4. tRAS Test


5. Command Rate Test


6. tRC Test


7. tRFC Test


8. tREF Test


Like Super PI test and unlike other ram timings, bigger tREF value is better at performance.

Let's see overall analysis.

Overall Analysis: Read Bandwidth


Overall Analysis: Write Bandwidth


Overall Analysis: Copy Bandwidth


In read bandwidth test, 4 most-affective ram-timing items are flatted around 2%. Command Rate-tREF-CL-tRFC are said affective items and other 4 ram-timings merely affect performance. But write bandwidth remains nearly unchanged during various ram-timing values has changed. In case of copy bandwidth, only tREF and tRFC exceed the affectiveness of 1% and others remain below 1%. In other words, only tREF and tRFC are actually affect copy bandwidth.

Then let's go to latency test.

- Latency test -

1. CL Test


In latency test, change of CL value derives nearly a linear trendine.

2. tRCD Test


Seems merely changed, but it's not 'that small' compared to latter results.
...At least the trendline is undoubtedly linear.

3. tRP Test


Little fluctuation is appeared, but it seems random empirical error.

4. tRAS Test


5. Command Rate Test


6. tRC Test


Guess there's only superior advantage on the value "tRC 25"? I guess not.

7. tRFC Test


8. tREF Test


Overall Analysis: Latency


Latency test seems more dependant to ram-timing values than two previous tests (Super PI & Bandwidth). 3 items exceed 1% of affectiveness and among them, 2 exceed 2.5% to performance. 2.5% performance increment upon no CPU / RAM clock change is not just a small variation.

Then let's go to the last test: video transcoding.


Video Transcoding test

This test is to measure the time of converting 1920x1080 TP file to MKV, same resolution. H.264 codec is used. No GPGPU acceleration.

1. CL Test


2. tRCD Test


3. tRP Test


4. tRAS Test


5. Command Rate Test


6. tRC Test


7. tRFC Test


8. tREF Test


Overall Analysis: Video Transcoding time


Similar to Super PI test, no items reached 1% and tREF marks the top. below CL, affectivenesses cannot even reach one third percent. But still remember: nearly 1% reduction in transcoding time with CPU & RAM clock remain unchanged is not 'that' small.


So, by far, we covered 3 areas: Pi calculation, Synthetic bench, and Video transcoding. Let's see how a ram-timing's affectiveness differ through that benchmark areas.

Affectiveness among test areas


As you see, tRP/tRAS and tRC have nearly "no" impact on performance.
Then let's calculate average affectiveness:
Average = {Super PI + (Read + Write + Copy)/3 + Latency + Encoding}/4
(name of each term indicate a ram-timing's affectiveness on that test area)

Average affectiveness by Ram-timing


最后我们看到对性能影响最大的是内存的CL值,其次是CR(1T/2T),这两项影响在1%以上,接下来第三是tRFC,也有接近1%的影响,然后是tREF,7.8ms性能反而好过3.9ms,接下来是对内存频率影响很大的tRCD,其实对性能影响只有0.5%左右。最后是tRP、tRC、tRAS对性能影响都很小,但是注意它们也同样会影响稳定性,尤其是tRAS,设太高太低都不好。
大D来了 发表于 2010-12-11 18:34 | 显示全部楼层
学习了,高频高时序最实惠...
liwenchao 发表于 2010-12-12 23:28 | 显示全部楼层
在频率不能高的时候就只能调时序了。
ssyknuwyg 发表于 2010-12-13 08:40 | 显示全部楼层
俺弄条宇瞻普条拉到1666,设置1t,其他全auto,调来调去太麻烦
playcent 发表于 2010-12-13 11:47 | 显示全部楼层
看来的确是高频高参最实惠~
Jeffryxu 发表于 2010-12-13 18:38 | 显示全部楼层
回复 1# royalk


    请问下R版 如果MEMTEST都过了 但是游戏出现蓝屏 又不想降频

    调时序放宽哪项比较好?
royalk  楼主| 发表于 2010-12-13 18:44 | 显示全部楼层
回复  royalk


    请问下R版 如果MEMTEST都过了 但是游戏出现蓝屏 又不想降频

    调时序放宽哪项比较 ...
Jeffryxu 发表于 2010-12-13 18:38


确定是内存的事?
Jeffryxu 发表于 2010-12-13 18:53 | 显示全部楼层
应该是吧 那对挫条 你知道的 代码忘了是多少了
royalk  楼主| 发表于 2010-12-13 18:56 | 显示全部楼层
应该是吧 那对挫条 你知道的 代码忘了是多少了
Jeffryxu 发表于 2010-12-13 18:53


其实主板插槽的事也不是没可能的。。如果没换过插槽可以试试
Jeffryxu 发表于 2010-12-13 19:01 | 显示全部楼层
这么挫啊? 但是杯具的是  冰刃一装  1 2号槽 直接在冰刃胯下了{:4_142:}

目前把TRCD放宽了一级
icz88 发表于 2010-12-15 17:27 | 显示全部楼层
全部参数加起来也没有5点。看来内存还是高频率划算。
xiaofeng000 发表于 2011-8-24 15:36 | 显示全部楼层
好文 学习了
慢慢研究
81894225 发表于 2011-8-27 11:54 | 显示全部楼层
之前写过一些文章提到内存超频只是辅助CPU超频,单独超内存本身并不会得到太多性能提升,但是却没有拿出实 ...
royalk 发表于 2010-12-11 18:05

R爷,能否给小弟 解释一下不太常调的那几个参数的意义呢 ?为何有的值越大越好?
royalk  楼主| 发表于 2011-8-27 12:04 | 显示全部楼层
R爷,能否给小弟 解释一下不太常调的那几个参数的意义呢 ?为何有的值越大越好? ...
81894225 发表于 2011-8-27 11:54


那几个值我也不知道,除非是学DRAM工作原理的
我也只会按着经验调
81894225 发表于 2011-8-27 12:19 | 显示全部楼层
那几个值我也不知道,除非是学DRAM工作原理的
我也只会按着经验调
royalk 发表于 2011-8-27 12:04

呵呵,可惜小弟虽然是学微电子的,也不知道这东西,惭愧啊。
royalk  楼主| 发表于 2011-8-27 12:22 | 显示全部楼层
呵呵,可惜小弟虽然是学微电子的,也不知道这东西,惭愧啊。
81894225 发表于 2011-8-27 12:19


这个其实可以从一些DRAM厂商的PDF上找到,感兴趣可以去看看。。涉及到电信号的那些东西,我反正是看不懂,你有这方面的知识应该还是能看懂的
81894225 发表于 2011-8-27 12:50 | 显示全部楼层
这个其实可以从一些DRAM厂商的PDF上找到,感兴趣可以去看看。。涉及到电信号的那些东西,我反正是看不懂 ...
royalk 发表于 2011-8-27 12:22

恩,还想问下像640sp的5670这个显卡的核心电压多少以下安全呢?(我的默认好像是1.125V)我的显卡不超频满载夏天大概65°,不知道这个算不算热的?还有对于非公版的显卡超频是不是只能刷bios了?(CCC就不考虑了),迪兰有没有自己显卡的超频工具?
royalk  楼主| 发表于 2011-8-27 12:52 | 显示全部楼层
恩,还想问下像640sp的5670这个显卡的核心电压多少以下安全呢?(我的默认好像是1.125V)我的显卡不超频满 ...
81894225 发表于 2011-8-27 12:50


1.125不高吧,5770也是这电压的,5870还要更高,65度还行了,没太大问题
不建议刷BIOS,CCC其实挺好,afterburner也可以开机应用超频设置
迪兰工具应该是有,但不知道做得怎样,没用过
81894225 发表于 2011-8-27 13:03 | 显示全部楼层
1.125不高吧,5770也是这电压的,5870还要更高,65度还行了,没太大问题
不建议刷BIOS,CCC其实挺好,aft ...
royalk 发表于 2011-8-27 12:52

哦,那就不刷BIOS了,ccc超频的那个我只能超25MHz稳定,高了就不行了,迪兰出厂已经超过频率了,850的核心频率,很坑爹。呵呵。

还想问下R大:如何判断主板的内存插槽的好坏呢?
royalk  楼主| 发表于 2011-8-27 13:29 | 显示全部楼层
哦,那就不刷BIOS了,ccc超频的那个我只能超25MHz稳定,高了就不行了,迪兰出厂已经超过频率了,850的核心 ...
81894225 发表于 2011-8-27 13:03


内存插槽?那得上到内存的最高频率才看得出来的
您需要登录后才可以回帖 登录 | 注册

本版积分规则

快速回复 返回顶部