WARNING Do not upgrade to 375/376.xx drivers (for xx<48)

[toTOW]

It's not the same issue ... th issue described in the thread here is an immediate failure when the core starts.

What you're seeing is GPU instabilities that can occur for whatever reason while running ... it's usually the Windows TDR feature triggering ...

[Applefreak]

I've just installed the latest Cuda 8 Dev package out of curiosity (1.2 GB download). It includes the 369.30 driver (7/14/2016). I believe this to be the latest stable release from Nvidia as it is meant for developers not gamers. Before, I had almost every WU return as bad on my GTX 960, running Windows 10. I am now back at folding without any issues so far.

[silverpulser]

I've just installed the latest Cuda 8 Dev package out of curiosity (1.2 GB download). It includes the 369.30 driver (7/14/2016). I believe this to be the latest stable release from Nvidia as it is meant for developers not gamers. Before, I had almost every WU return as bad on my GTX 960, running Windows 10. I am now back at folding without any issues so far.

I have reverted back to 373.06 and that too is very stable and returned my points to normal. I am using an NVidia GTX 750Ti. Also now on FAH v7.4.16 without any problems so far (only installed the update Monday).

[Skram0]

Driver 376.48 has been working great for me 24x7 since 1/5/17 on my ASUS Strix GTX 1060. This was a new video card install, upgraded from an aging AMD, and I thought I'd download the latest video driver. But after seeing my graphics crash in GPU TweakII and FAH instantly crashed, I decided to do some searching and found numerous threads relating to the latest drivers. So I installed the hotfix and it seems to work okay. Doom plays good too.

[Sven]

Seen in viewtopic.php?f=24&t=29556 , Anton writes about a updated CUDA-Version.

Isn't Core21 a OpenCL Application for Nvidia and AMD?

[foldy]

Maybe they mean an updated CUDA SDK 8.0 which provides the OpenCL SDK? Or Nvidia has some money to spend and they push a CUDA version again which solves the OpenMM issue with nvidia OpenCL driver by not using OpenCL anymore

[bruce]

In theory, the whole point of OpenCL is to provide a common interface that works with anybody's hardware. FAHCore_21 seems to work fine with the OpenCL v1.2 provided by AMD and, since the release of 375.xx, does not work well with OpenCL v1.2 provided by NV. Somewhere there's a hole in the OpenCL theory.

In the past, FAH has sometimes created a pair of Cores to process the same WUs; one using OpenCL and another using CUDA. (CUDA is NV proprietary so AMD doesn't support it. Officially both support OpenCL.) There are added costs to develop, test, and maintain two versions when one will do. They seem to have decided that the advantages provided by two is worth the extra costs in order to circumvent the problems discussed throughout this topic.

We'll see when it happens.

[7im]

Additional data validation costs running 2 sets of code as well, making sure NV and AMD return the same results.

[Aurum]

I've just installed the latest Cuda 8 Dev package out of curiosity (1.2 GB download). It includes the 369.30 driver (7/14/2016). I believe this to be the latest stable release from Nvidia as it is meant for developers not gamers. Before, I had almost every WU return as bad on my GTX 960, running Windows 10. I am now back at folding without any issues so far.

Thanks for the tip. I was using the 376.48 hotfix but my rig PPD was about 1,500,000. I then did an uninstall/clean and ran 372.70 but PPD only rose to 1,700,000. Yesterday I got 369.30 and my PPD has hit 2,030,000. It's just been a day but that's the highest I've ever seen it.
Note: Even though the GTX 970 is on a x1 slot it almost always has the highest Estimated PPD. Is Maxwell better at folding than Pascal

11:42:05: CPU: AMD FX(tm)-8350 Eight-Core Processor
11:42:05: CPU ID: AuthenticAMD Family 21 Model 2 Stepping 0
11:42:05: CPUs: 8
11:42:05: Memory: 16.00GiB
11:42:05: Free Memory: 13.76GiB
11:42:05: Threads: WINDOWS_THREADS
11:42:05: OS: Windows 7 Professional Service Pack 1
11:42:05: OS Arch: AMD64
11:42:05: GPUs: 4
11:42:05: GPU 0: Bus:6 Slot:0 NVIDIA:5 GM204 [GeForce GTX 970] (Bus runs at x1)
11:42:05: GPU 1: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x8)
11:42:05: GPU 2: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x8)
11:42:05: GPU 3: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x16)

[Slowfood]

Does someone has an dirver recommondation for linux x64 folding with an GTX 1050 Ti? Nvidia supported them the first time with the 375.10 Linux driver

[SombraGuerrero]

I tend to air on the conservative side, but honestly, if you're able to fold with the driver you currently have, I wouldn't mess with it. If you're experiencing the bug with 0x21 though, I would recommend pinning your kernel and installing the hotfix driver if you're comfortable with manual driver installation. In my experience, pinning the kernel is necessary, especially if you want to keep your system stable, because every time the kernel upgrades, manually installed Nvidia drivers tend to break.

[JT3rd]

Does someone has an dirver recommondation for linux x64 folding with an GTX 1050 Ti? Nvidia supported them the first time with the 375.10 Linux driver

I tend to air on the conservative side, but honestly, if you're able to fold with the driver you currently have, I wouldn't mess with it. If you're experiencing the bug with 0x21 though, I would recommend pinning your kernel and installing the hotfix driver if you're comfortable with manual driver installation. In my experience, pinning the kernel is necessary, especially if you want to keep your system stable, because every time the kernel upgrades, manually installed Nvidia drivers tend to break.

Thanks for the advise, I plan on moving back over to Linux Mint XFCE in a few weeks, just waiting to pick up a separate SSD so I can utilize the full encryption without jeopardizing my current windows build. I'm curious to see how the 1050/1050Ti will perform on the Linux side.

[SombraGuerrero]

Okay, the important thing is that the device code for the 1050 Ti is 1C82 (the 1050 is 1C81).
Open nv_dispi.inf from 373.06 and any 375/6 version in your favourite text editor. Find a line similar to
"%NVIDIA_DEV.1C82% = Section133, PCI\VEN_10DE&DEV_1C82 " in the newer file then copy and paste it into the older file after each instance of a similar line with 1C03 which is the code for the 6GB GTX 1060.
Change the Section number to match that of the 1C03 line each time.
Finally, there's a line near the end of the file that maps device ID to device name so copy that across too.
Save the modified file and install 373.06.

This was also done by a user earlier in this thread as a crafty workaround. I don't know how well, if at all, this process would translate to the Linux environment, and most certainly this would probably be some sort of .conf or xml file, maybe a shell script, but I thought I'd throw this out there as a different idea.

[rwh202]

Note: Even though the GTX 970 is on a x1 slot it almost always has the highest Estimated PPD. Is Maxwell better at folding than Pascal
11:42:05: GPU 0: Bus:6 Slot:0 NVIDIA:5 GM204 [GeForce GTX 970] (Bus runs at x1)
11:42:05: GPU 1: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x8)
11:42:05: GPU 2: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x8)
11:42:05: GPU 3: Bus:4 Slot:0 NVIDIA:5 GP104 [GeForce GTX 1070] (Bus runs at x16)

More likely is that the slot numbering / naming is wrong. GTX 970 will be ~300k and 1070 at ~550k. What happens if you start each GPU slot up one at a time and see which one loads and heats up? The latest beta clients are meant to be an improvement in this regard, but the previous client was rarely (never?) correct.

[Slowfood]

I tend to air on the conservative side, but honestly, if you're able to fold with the driver you currently have, I wouldn't mess with it. If you're experiencing the bug with 0x21 though, I would recommend pinning your kernel and installing the hotfix driver if you're comfortable with manual driver installation. In my experience, pinning the kernel is necessary, especially if you want to keep your system stable, because every time the kernel upgrades, manually installed Nvidia drivers tend to break.

I'm asking, because i experience the 0x21 bug. I'm using the NVIDIA-Linux-x86_64-375.26 driver. Manually installed. The 1050 Ti is offically supported since the 375.10

Code: Select all

23:26:58:WU01:FS01:FahCore 0x21 started
23:26:58:WU01:FS01:0x21:*********************** Log Started 2017-01-21T23:26:58Z ***********************
23:26:58:WU01:FS01:0x21:Project: 9191 (Run 1, Clone 16, Gen 247)
23:26:58:WU01:FS01:0x21:Unit: 0x00000166ab40415457cb2cb4df8ee9c7
23:26:58:WU01:FS01:0x21:CPU: 0x00000000000000000000000000000000
23:26:58:WU01:FS01:0x21:Machine: 1
23:26:58:WU01:FS01:0x21:Reading tar file core.xml
23:26:58:WU01:FS01:0x21:Reading tar file system.xml
23:26:58:WU01:FS01:0x21:Reading tar file integrator.xml
23:26:58:WU01:FS01:0x21:Reading tar file state.xml
23:26:58:WU01:FS01:0x21:Digital signatures verified
23:26:58:WU01:FS01:0x21:Folding@home GPU Core21 Folding@home Core
23:26:58:WU01:FS01:0x21:Version 0.0.17
23:27:01:WU01:FS01:0x21:ERROR:exception: Error downloading array interactionCount: clEnqueueReadBuffer (-5)
23:27:01:WU01:FS01:0x21:Saving result file logfile_01.txt
23:27:01:WU01:FS01:0x21:Saving result file log.txt
23:27:01:WU01:FS01:0x21:Folding@home Core Shutdown: BAD_WORK_UNIT
\x1b[93m23:27:07:WARNING:WU01:FS01:FahCore returned: BAD_WORK_UNIT (114 = 0x72)\x1b[0m
23:27:07:WU01:FS01:Sending unit results: id:01 state:SEND error:FAULTY project:9191 run:1 clone:16 gen:247 core:0x21 unit:0x00000166ab40415457cb2cb4df8ee9c7
23:27:07:WU01:FS01:Uploading 2.52KiB to 171.64.65.84
23:27:07:WU01:FS01:Connecting to 171.64.65.84:8080
23:27:07:WU00:FS01:Connecting to 171.67.108.45:80
23:27:08:WU01:FS01:Upload complete
23:27:08:WU01:FS01:Server responded WORK_ACK (400)
23:27:08:WU01:FS01:Cleaning up