=======Core21 v0.0.17 has been released!========

金鹏

在接到114XX和1170X及91XX的21内核包时会被强制更新到0.017内核

关闭PME,性能继续下降8.5+%,回归到和0.014一个水平尿性
980TI跑11410从0.016的73万跌倒0.017的62万,损失11万PPD而1380下980TI跑11411包在0.017内核下自有58万,在0.016内核下有70万
基本上0.017内核下比0.016内核下低了20+%的PPD






FahCore_21 Version 0.0.17.rar (2.43 MB, 下载次数: 13)

2015-12-22 19:07 上传

点击文件名下载附件
阅读权限: 10

Core21 v0.0.17 has been released!

by JohnChodera » Thu Dec 31, 2015 3:17 am

We rolled out Core21 v0.0.17 earlier this week.

Summary of changes:
* Checkpoint sanity checks are now multithreaded (when supported by hardware). When possible, the multithreaded OpenMM CPU platform is now used for sanity checks prior to checkpointing instead of the much slower single-threaded Reference platform. This should reduce the impact of these sanity checks on performance, and allow Core 21 projects to run better alongside multithreaded CPU cores.
* Core logs now print frames/WU and checkpoint interval. This is currently experimental, but will eventually be added to the client logs.
* Warning/reminder about overclocking now issued after Bad State failures.

We are still working on a more permanent fix for the performance regression for PME on NVIDIA introduced in 0.0.14 to greatly improve stability.

Thanks to everyone who helped make this possible, especially Peter Eastman and Joseph Coffland, as well as everyone who helped test:
CODE: SELECT ALLhayesk
frodothehobbit
davidcoton
muziqaz
grandpa
totow
christianvirtual
bollix47
jimerickson


Happy New Year, all!

~ The Chodera and Pande groups

New Core21 build 0.0.17 live in BETA!

Moderators: Site Moderators, PandeGroup

First unread post • 3 posts • Page 1 of 1

New Core21 build 0.0.17 live in BETA!

[size=1em]by JohnChodera » Tue Dec 22, 2015 4:21 am

core21 v0.0.17 has been rolled out to BETA. We'll be updating the minimum core versions on BETA projects over the next few days, so please bear with us.

CHANGELOG:
* Revert back toCODE: SELECT ALLusePmeQueue = falseto reduce frequency of Bad State errors with NVIDIA cards, again incurring ~8.5% performance regression

We're planning to use the holidays to study the Bad State issue on NVIDIA hardware and see if we can come up with a fix that doesn't cause cause performance loss.

eqzero

最近core21好像升级很频繁啊~

金鹏

今天貌似916X的强制升级21内核的毒包没了，
可以在beta目录下把0.017的换回0.016的跑114XX和1049X的21内核包

金鹏

Core21 v0.0.17 has been released!

by JohnChodera » Thu Dec 31, 2015 3:17 am

We rolled out Core21 v0.0.17 earlier this week.

Summary of changes:
* Checkpoint sanity checks are now multithreaded (when supported by hardware). When possible, the multithreaded OpenMM CPU platform is now used for sanity checks prior to checkpointing instead of the much slower single-threaded Reference platform. This should reduce the impact of these sanity checks on performance, and allow Core 21 projects to run better alongside multithreaded CPU cores.
* Core logs now print frames/WU and checkpoint interval. This is currently experimental, but will eventually be added to the client logs.
* Warning/reminder about overclocking now issued after Bad State failures.

We are still working on a more permanent fix for the performance regression for PME on NVIDIA introduced in 0.0.14 to greatly improve stability.

Thanks to everyone who helped make this possible, especially Peter Eastman and Joseph Coffland, as well as everyone who helped test:
CODE: SELECT ALLhayesk
frodothehobbit
davidcoton
muziqaz
grandpa
totow
christianvirtual
bollix47
jimerickson


Happy New Year, all!

~ The Chodera and Pande groups