new core A5 projects: 8106-8108启动/Big Adv Program Ends January 31

金鹏

Re: Updates thread

by kasson » Thu Feb 05, 2015 3:01 am

Projects 8106-8108 have been pulled from assign. We're re-calculating points yields. There was a mistake in points assignment that also triggered a safety feature in the stats code. We're fixing all these problems and will be *slowly* re-releasing through the beta process as everything comes online. Many apologies for the problems here.

金鹏

Re: new core A5 projects: 8106-8108

by kasson » Thu Feb 05, 2015 3:09 am

Many apologies for the problems here. I was trying to ensure work unit availability for a smoother transition post-bigadv; in retrospect, it would have been better to accept a "gap" period for many-core-optimized work units and make sure that the rollout process worked properly.
What happened here was the following:
1. In benchmarking the work units, I mis-combined formulae and ended up with a k-factor that was substantially too high.
2. The very high bonus values that resulted triggered a safety threshold in the stats code, resulting in no bonus points at all for some work units.
3. We had a hardware failure on one of the benchmarking machines, slowing correction of this process.

We are in the process of fixing all three of these. Once these are done, we will proceed with a slow beta rollout. Be advised that PPD values will be lower--they were inflated substantially by the benchmarking mistake I made. Many apologies.

金鹏

Re: new core A5 projects: 8106-8108

by kasson » Thu Feb 05, 2015 9:37 am

The missing-bonus returns have now been rerun and should have appropriate credit.

Lynt

这是要减分的节奏？

wpf999

金鹏 发表于 2015-2-5 08:50

斯坦福大学的研究员也能代错公式，这不科学。。。

金鹏

wpf999 发表于 2015-2-5 13:17
斯坦福大学的研究员也能代错公式，这不科学。。。

跑了一天的7504小包,才13万PPD,跑100分钟一个包

Re: new core A5 projects: 8106-8108

by kasson » Thu Feb 05, 2015 1:49 am

We've shut off assigns for these projects until we can resolve problems. Thanks for your patience.

PS we have records for all returned work units.

Re: new core A5 projects: 8106-8108

by kasson » Thu Feb 05, 2015 9:37 am

The missing-bonus returns have now been rerun and should have appropriate credit.

金鹏

Lynt 发表于 2015-2-5 11:00
这是要减分的节奏？

K因数大幅缩水

Re: new core A5 projects: 8106-8108

by kasson » Sat Feb 07, 2015 11:35 am

We are getting ready to restart project 8106 on beta. k-factor is 3.92; all other numbers are unchanged. 8107 and 8108 will follow in the next several days as our benchmark re-runs finish.

The changes at this time are only in project availability, no core changes.

wpf999

金鹏 发表于 2015-2-7 20:19
K因数大幅缩水

双2678w现在多少PPD？

金鹏

wpf999 发表于 2015-2-7 20:23
双2678w现在多少PPD？

加了bigbeta参数在接8106新包,下载速度坑,大小从原来的23兆变成现在29兆+

Re: new core A5 projects: 8106-8108

by EXT64 » Sat Feb 07, 2015 1:03 pm

That PPD looks much more in line with standard SMP. A couple comparisons:

8106 -> 130k PPD
8822 -> 139k PPD
7504 -> 120k PPD
8814 -> 133k PPD

So I would say benchmarking was a success.

Edit: for reference this is a 2P E5-2670.

cuda

金鹏 发表于 2015-2-7 20:19
K因数大幅缩水

8106原先的k-factor是39.2，现在改成3.92，之前是点错了小数点？
由于PPD和k-factor的1/2方成正比，那么8106的PPD应该缩水为原先的1/sqrt(10)=0.316倍。双路E5-2699v3的能效比也不值得推荐了。

金鹏

cuda 发表于 2015-2-7 20:57
8106原先的k-factor是39.2，现在改成3.92，之前是点错了小数点？
由于PPD和k-factor的1/2方成正比，那么8 ...

兄弟精确43*.316=13万
我现在32用了30T跑的12.8万PPD

金鹏

Re: Updates thread

by kasson » Thu Feb 12, 2015 12:51 am

Projects 8106 and 8108 are now on beta once more, with recalculated points and k-factors. According to reports thus far, the PPD is roughly in line with other A3 work units that run well on many cores.

wpf999

金鹏 发表于 2015-2-12 16:15

fah官方其实是为了鼓励大家用GPU，又快又省电

金鹏

wpf999 发表于 2015-2-13 13:10
fah官方其实是为了鼓励大家用GPU，又快又省电

32T也放弃了,全面转跑WCG刷牌子,未来就留O版捐助的64C跑鸡血下8106-8108

Re: new core A5 projects: 8106-8108

by kasson » Sat Feb 14, 2015 2:24 am

Thanks for the reports thus far. 8107 is also up:
Points: 7358.0
Preferred deadline: 4 days
Final deadline: 54 days
k-factor: 3.92

yimu35

5960X跑这个FAH还是WCG好？