POPPUR爱换

标题: 关于诸位高人Super Pi 1M的速度分析[以及关于Coolaler和OPB OC作弊可能性的分析] [打印本页]

作者: EnergyStar 时间: 2006-12-15 16:53
标题: 关于诸位高人Super Pi 1M的速度分析[以及关于Coolaler和OPB OC作弊可能性的分析]
http://www.xtremesystems.org/forums/showthread.php?t=126560

First off, in light of TaPaKaH's recent admittance of faking everything, we're not scheming to get anyone, or anything like that. He did on his own. However, let's put that in perspective as well. We were all fooled for MONTHS. That's how possible, legitimate, and prevalent cheating really is....and it's a shame.

至于这个TaPaKaH呢...就是当初跑E6300/ E6400 WR的人.

In zbogorgon's recent thread: http://www.xtremesystems.org/forums/showthread.php?t=125972, he further elucidates how cheating can be done easily....he's due a lot of credit for coming forward and bringing this up, as well as recognizing where some cheating may exist.

Let's remember one key thing here....the methods that we know exist, speedhack linearly.

Well, SPi doesn't scale linearly, it's inversely proportional. You can't linearly generalize: 100MHz increase is equal to a given time decrease. However, you can say this: MHz * time ~= X, where X is constant for any given setup.

For a high speed, VERY well tweaked 975X + Conroe sytem, X is equal to about 50500. Remember this.

Let's take a look at some top SPI 1.5 results, taken from hwbot.org and from on the forum.

OPB - 5550 x 8.91 = 49451 (Oct 16) here
OPB - 5567 x 9.16 = 50994 (Oct 10) **
Maxxxpct - 5506 x 9.33 = 51370
Coolaler - 5525 x 9.00 = 49725 here
TaPaKaH - 5408 x 9.36 = 50619
OPB - 5407 x 9.39 = 50772 (Sept 24) **
Memoryextreme Team - 5348 x 9.44 = 50485
Hipro5 - 5376 x 9.45 = 50803
OPB - 5329 x 9.55 = 50892 (Sept 8) **
ojdr2001 - 5348 x 9.55 = 51073
OPB - 5305 x 9.63 = 51087 (Sept 2) **
KP - 5297 x 9.68 = 51275
Hipro - 5217 x 9.72 = 50709
TaPaKaH - 5208 x 9.78 = 50934
zbogorgon - 5196 x 9.8 = 50920
hipro - 5138 x 9.89 = 50815
Kyosen - 4167 x 12.05 = 50212 here. FWIW, this is THE BEST clock-for-clock time to be found...that's legit at least
这个是目前非A/C的air 1M WR，也是目前已知最好的优化（排除可能作弊的OPB和Coolaler）
before - 5000 x 10.06 = 50300 here...look at that RAM speed!
kyosen - 9.937 x 5086 = 50540 here

See a trend? Known, reputable SuperPi GODS need lowish MHz, insane tweaks, and insane RAM to barely break below 50400.

Let's now take a look at Coolaler's progression, from slowest to fastest:

5200 x 9.72 = 50544 (July 12)
5300 x 9.64 = 51092 (July 17)
5396 x 9.485 = 51181 (July 18)
5409 x 9.484 = 51299 (July 19)
Now things begin to get funky
5485 x 9.204 = 50484 (July 23)
5512 x 9.156 = 50468 (July 28)
Now the grand finale:
5525 x 9.000 = 49725 (Aug 2)

Not a single one of those final three has outstanding FSB/RAM speed or timings, not a single one has much headroom for MHz suicide stability....and are EXCEEDINGLY fast, especially considering his tweaking skills. Something changed....something we can't see.

Well, let's take a look at his results from above....see clock-for-clock improving over time? Then he drops the bomb: 8.906 at 5550. He also later shares all his tweaks, they're old, they don't do that much (if anything) on C2D, people have tried them. Then, if you also look back at his non-WR results, you see this gem:

4710 x 10.344 = 48720 here

最荒谬的成绩出现...

....which is just crazy fast, fast to the point where he probably ran it at a higher speed than what he had in the screenshot.

Anyway, something's up with the orange and red results....they're just not legit. The green results are the best we could find....and are by known SPI legends, and have the hardware setting to backup the clock-for-clock greatness they have.

Looking at the top 250 results between hwbot.org and Coolaler and OPB's posts, and charting them in excel....well, pictures certainly can say a thousand words.

Quick explanation: each point represents a SPI run, look at the axis to tell what each graph is graphing....the "average" line is, based on the top 250 results and their clock-for-clock performance, that is the "average expected performance line". The "ideal" line is based on "50500"....a.k.a., a VERY well-tweaked system, and what you can expect as MHz changes.

Coolalers and OPB's results have been singled out so you can tell which are theirs. Picoo's recent run is also singled out as the new WR, and to show, that what is expected from an insanely tuned system from one of Team Japan's members.

Some sort of explanation is due....either it's a bug, a cheat, or a tweak that does what no other tweak can do in 1M: improve times by more than a few hundreths of a second.

以上内容相信各位的E文水平也能看懂，没有什么好解释的了

放几个隔壁的成绩对比下好了
VictorWang 10.969*4659.6=50971
肥婆 11.047*4644.1=51303
VC 11.875*4300=51063
K.J. 11.937*4238.5=50595
肉佬 12.984*3903.4=50681

肉佬用的是B1，如果说跑pi的话，相同的tweaking，B2平均会比B1快0.046-0.047s左右
(12.984-0.047)*3903.4=50498

这是目前内地唯一已知有可能进50500的人

Coolaler在两年前的3dmark测试中也有作弊嫌疑，具体见我最上面引用的XS的帖子的具体分析

[ 本帖最后由 EnergyStaR 于 2006-12-15 21:45 编辑 ]

作者: sammy2005 时间: 2006-12-15 17:07
提示: 作者被禁止或删除内容自动屏蔽

作者: zaarath 时间: 2006-12-15 17:11
这个Hz*Second在同样设置的系统中基本保持一致的特性我早就注意过，每次自己跑完Pi都自己算一下。

但是这个分析并不能完全说明问题。
首先OPB/Coolaler创造纪录时所用的FSB并不高，都在400左右，这样他们可以通过对内存狂加电压来达到极低的Timing，这样Hz*Second就会很低，这需要opb/coolaler报出timing才能说明问题。

第二点跑spi有很大的偶然性，大家肯定有过这样的经验，你跑出一个最好成绩以后，可能再也达不到了，这个是运气问题。

最后spi得计时并不是连续的，而是会跳跃的，比如你达到9.00, 下一个成绩可能就是8.90，无论跑多少次也达不到8.95。

怀疑可以，但是统计学的基本原理就是无法证明小概率事件的不可能性，只要差距不是太离谱，这个证据是不够的。

作者: EnergyStar 时间: 2006-12-15 17:22

原帖由 zaarath 于 2006/12/15 17:11 发表
这个Hz*Second在同样设置的系统中基本保持一致的特性我早就注意过，每次自己跑完Pi都自己算一下。

但是这个分析并不能完全说明问题。
首先OPB/Coolaler创造纪录时所用的FSB并不高，都在400左右，这样他们可 ...

timing见他们的截图就有， GMH/GKX无论怎么加电压在800左右或以上也只能在3-3-3 重要的是频率并不高

要离谱的成绩么，看OPB那个4.71g的成绩好了

至于你说的计时连续问题，那是因为有些值根本不会取到，spi1.5有个取时精度问题

作者: zaarath 时间: 2006-12-15 17:30

原帖由 EnergyStaR 于 2006-12-15 17:22 发表

timing见他们的截图就有， GMH/GKX无论怎么加电压在800左右或以上也只能在3-3-3 重要的是频率并不高

要离谱的成绩么，看OPB那个4.71g的成绩好了

至于你说的计时连续问题，那是因为有些值根本不会取到， ...

我不玩极限OC所以对tweaking也不是很懂，只是根据自己的经验来说。正是因为有些值不会取到所以spi的显示有可能比真正用时要高。

举例来说spi的报时跳跃是0.1秒（仅仅是假设）
比如一个人真实时间是 8.951秒，spi可能就只能报出9.00
另一个人真实时间是8.949秒，spi可能报出8.90的成绩
这样几乎一样的成绩却造成0.1秒的巨大差异。

作者: Dr.BT 时间: 2006-12-15 17:35
ES，数字错了
是50500
上点统计数据
看最后一项吧

作者: VictorWang 时间: 2006-12-15 19:49

原帖由 EnergyStaR 于 2006-12-15 17:22 发表

timing见他们的截图就有， GMH/GKX无论怎么加电压在800左右或以上也只能在3-3-3 重要的是频率并不高

要离谱的成绩么，看OPB那个4.71g的成绩好了

至于你说的计时连续问题，那是因为有些值根本不会取到， ...

:unsure: 要不要看ddr800 3-3-2-8 ?

不要别人有好成绩就跟在后面酸，没意思的啦，老外就是放假在家无聊，把台湾朋友搞了一圈，最后轮到opb而已。
opb对oc的执著，对tweak的研究是你们无法想象的，所以呢，任何事情不要妄加评论。

作者: VictorWang 时间: 2006-12-15 19:51
要看成绩，我这里有一个，低于50500，是不是作弊阿？

作者: hopetoknow 时间: 2006-12-15 20:57
老外真落伍。咱N久前，就知道算主频*时间啦

是比较惊人

作者: ylm8445 时间: 2006-12-15 21:34
都是高手,我看不懂
估计猴子看的懂

欢迎光临 POPPUR爱换 (https://we.poppur.com/)