PC Drivers & Software

[Dennis_Menace]

We are expiring errors (maxErr) on the AMDOCL.dll OCL2 transform driver from the driver packages newer than 3240.6 from the Radeon Software 21.5.2.
These errors have been confirmed (until now) on the Radeon RX6700XT and RX6600XT cards running GNF-15 to GNF-19 tasks on Boinc from PrimeGrid.

GNF-20 to 22 that uses OCL and all other tasks from Boinc - PrimeGrid run with no issues.
Following is the Stderr output from a GNF-16 task attempt on my RX6700XT

Stderr output

<core_client_version>7.16.11</core_client_version>
<![CDATA[
<message>
The system cannot find the device specified.
 (0x14) - exit code 20 (0x14)</message>
<stderr_txt>
geneferocl 3.3.3-2 (Windows/OpenCL/32-bit)

Copyright 2001-2018, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2014, Michael Goetz, Ronald Schneider
Copyright 2011-2018, Iain Bethune
Genefer is free source code, under the MIT license.

Running on platform 'AMD Accelerated Parallel Processing', device 'gfx1031', vendor 'Advanced Micro Devices, Inc.', version 'OpenCL 1.2 AMD-APP (3302.6)' and driver '3302.6 (PAL,LC)'.
20 computeUnits @ 2548MHz, memSize=3072MB, cacheSize=16kB, cacheLineSize=64B, localMemSize=64kB, maxWorkGroupSize=256.
Supported transform implementations: ocl ocl2 ocl3 ocl4 ocl5 

Command line: projects/www.primegrid.com/geneferocl_windows_3.3.3-2.exe -boinc -q 134332300^65536+1 

Normal priority change succeeded.
Checking available transform implementations...
OCL transform is past its b limit.
OCL3 transform is past its b limit.
OCL4 transform is past its b limit.
OCL5 transform is past its b limit.
Using OCL2 transform
Starting initialization...
Initialization complete (0.097 seconds).
Testing 134332300^65536+1...
Estimated time for 134332300^65536+1 is 0:01:18                 

maxErr exceeded for 134332300^65536+1, 1.0000 > 0.4500
Errors occurred for all available transform implementations
Waiting 10 minutes before attempting to continue from last checkpoint.. 

[Dennis_Menace]

Ohhh let me add that this error occur on Windows 10 platforms using all later drivers than 3240.6 from Radeon Software 21.5.2 that we checked on PrimeGrid

[Dennis_Menace]

It seems like you want a lot more info, so here goes:

Edition Windows 10 Enterprise
Version 21H1
Installed on ‎9/‎7/‎2021
OS build 19043.1288
Experience Windows Feature Experience Pack 120.2212.3920.0

ASUS Tuf Gaming X570-Plus motherboard - Bios Version 4021. date 2021/08/17
GPU Sapphire NITRO+ AMD Radeon™ RX 6700 XT - driver Radeon Software Adrenaline 21.10.2 
Ryzen 9 3900X CPU
Corsair HX1200W PSU
GSkill FlareX 3200Mhz 4x8GB Ram
Display Gigabyte G27QC

[hsuknowledge]

Based on PrimeGrid stats, hosts of RX 6900 XT (1), RX 6700 XT (2), and RX 6600 XT (1) cards, encountered the same arithmetic error. Their task results print out "maxErr exceeded, 1.00 > 0.45". Also, when testing the affected OpenCL program, a roundoff error "1.00 > 0.45" is emitted in very few iterations as I abort it, or eventually it will report the "maxErr exceeded" message before the iterations count down to zero.

The affected hosts report to have in common:

1. RX 6000 series card
2. "OpenCL 1.2 AMD-APP (3276.6) and driver 3276.6 (PAL,LC)" and the later driver (3302.6, as of now)
3. Windows 10

On the contrary, an RX 5700 XT host with the "3276.6 (PAL,LC)" driver reported to compute the tasks normally. It does not affect Polaris or Vega cards either, though I think their drivers are different.

The Linux driver for my RX 6600 XT, amdgpu-pro 21.30, which has "OpenCL 2.0  and driver 3314.0 (HSA1.1,LC)", also does not make the above roundoff error.

By the AMD Bug Report Tool I have attached the affected program in a report. I'm afraid I could not upload it to the open space.

[Dennis_Menace]

mail from AMD Community:

[quote]Hello Dennis_Menace,
Your topic recently received replies.

Topic: Error on RX6000 series cards AMDOCL.ddl drivers newer than 3240.6 from Radeon Software 21.5.2
Date: 10-14-2021 01:56 AM
Did it solve your problem?
Click here to view the replies and mark one as an Accepted Solution.
This helps others find helpful answers in the community too!

Thanks for being a AMD Community member.
Your AMD Community Team
AMD Community sent this message to "email adress".
Don't want to receive emails from us? Just click here.
[/quote]

No It is you (AMD) that need to fix this problem on the next RX6000 series drivers.. We just posted the problem to make you aware of that there is a problem on the RX6000 series GPU drivers in Windows 10 on the AMDOCL.dll drivers newer than the ones that come along after the Radeon software 21.5.2. That one driver and older worked well. (See the post)
The reply you say I received is another member confirming my findings (read the post)

I also replied to your mail in hope you will see it this time

Thanks in advance. We are looking forward to an updated (working driver) Dennis The Menace

[Dennis_Menace]

The error still persists on the new drivers coming along with the new Radeon Software Adrenalin 21.10.3. Released 25. Oktober 2021.
My Daughters older and smaller NVidia GTX1650 run all PrimeGrid tasks with NO issues! And the few tasks my RX6700XT can actually run. Her small GTX1650 run only a few seconds slower than my super expensive AMD RX6700XT!!!
AMD have neglected keeping their AMDOCL.dll drivers tuned and in working condition for their super expensive RX6000 series GPUs. It is embarrassing that my daughter can buy a NVidia GPU to 1/10 of the price of mine and can run all the tasks and to close to the same performance?
Just have a look at PrimeGrid fastest GPU list... https://www.primegrid.com/gpu_list.php 
AND GPUGRID on Boinc don't even bother making their tasks compatible with AMD GPUs because AMD neglect keeping AMDOCL.dll up to date and tuned.

It is this kind of things that makes it really difficult to keep supporting AMD and is making me regret spending so many money on a AMD RX6700RX
(Does anybody want to buy a slightly used Sapphire Radeon RX6700XT +Nitro so i can go back to NVidia)?

[Dennis_Menace · Accepted solution]

Problem solved on AMDs new "Adrenalin 21.12.1 Optional" software driver package released 12/3/2021