cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

sdanchenko
Journeyman III

Errors in memtestCL when testing more then 3992Mb

My setup:

Linux Ubuntu 12.10

Sapphire Radeon Vapor-X HD 7970 GHZ OC 6GB

AMD 13.2 driver and APP SDK 2.8

I would like to test memory on the card. If I run "memtestCL 3992 1" I have no errors. When I run with 3994 or more I got tons of errors.

Also, as a last resort, I inserted "cout << CL_DEVICE_MAX_MEM_ALLOC_SIZE <<endl;" before clCreateBuffer command and got printed out 4112 as Max allocatable memory (if I understand it correctly).

So, my question: is there a bug in memtestCL or I have, actually, memory errors and it is better to replace a card?

Also, are there any other ways to test GPU memory? Especially 6Gb.

Regards,

Serhiy.

0 Likes
1 Solution

That's correct. The driver you have supports 2GB for a single process. The limitation comes from the 32 bit address space currently used on the GPU. It may take some time until the 64 bit mode on HD7970(6GB) will be available in OpenCL by default.

View solution in original post

0 Likes
7 Replies
sdanchenko
Journeyman III

Update:

I set environment variable "set GPU_FORCE_64BIT_PTR=1"  according to http://devgurus.amd.com/thread/160325 post and was able to get to 4096Mb of no errors. Anything over 4096Mb still gives me tons of errors.

So, I kinda lost. Is it AMD driver issue? Is it hardware issue? Is it memtestCL issue?

Also, I am using 64bit Linux.

Regards,

Serhiy

0 Likes

sdanchenko wrote:

Update:

I set environment variable "set GPU_FORCE_64BIT_PTR=1"  according to http://devgurus.amd.com/thread/160325 post and was able to get to 4096Mb of no errors. Anything over 4096Mb still gives me tons of errors.

So, I kinda lost. Is it AMD driver issue? Is it hardware issue? Is it memtestCL issue?

"set GPU_FORCE_64BIT_PTR=1" is a key for a new feature, which was not intended for the end users. Currently with the latest drivers it may have problems with the generated opencl kernels.

You can test only memory reported by the OpenCL runtime with the default settings (see my message regarding clinfo).

Use that value(from "Global memory size", converted to MB) in "memtestCL". Also you can run a couple instances of "memtestCL" at the same time to test more memory on your system.

0 Likes
german
Staff

sdanchenko wrote:

Also, as a last resort, I inserted "cout << CL_DEVICE_MAX_MEM_ALLOC_SIZE <<endl;" before clCreateBuffer command and got printed out 4112 as Max allocatable memory (if I understand it correctly).

That's not correct. That line prints the value of CL_DEVICE_MAX_MEM_ALLOC_SIZE define.

Run "clinfo" program on your system(it's a part of the installation). It will print the device information, supported by the OpenCL runtime on your system(both CPU and GPU). Find the line with "Global memory size" for your GPU device.

For example:

            Global memory size:                            1073741824

That means 1GB.

0 Likes

Hi German,

Thank you for respond.

This is my clinfo output:

$ clinfo | grep mem

  Max memory allocation:             536870912

  Global memory size:                 2147483648

  Local memory type:                 Scratchpad

  Local memory size:                 32768

  Unified memory for Host and Device:         0

  Max memory allocation:             4198122496

  Global memory size:                 16792489984

  Local memory type:                 Global

  Local memory size:                 32768

  Unified memory for Host and Device:         1

Which does not make sense as it shows only 2147483648 / 1073741824 = 2 Gb. Most Radeon HD7970 comes with 3Gb and my card suppose to have 6Gb.

PS. Sorry for late replay - got sick and did not have access to computer.

0 Likes

That's correct. The driver you have supports 2GB for a single process. The limitation comes from the 32 bit address space currently used on the GPU. It may take some time until the 64 bit mode on HD7970(6GB) will be available in OpenCL by default.

0 Likes

IIRC this limitation is mentioned in OpenCL programing guide.

0 Likes

If there is a limitation of 2Gb per process. How can I test GPU memory besides starting of couple memtestCL in the same time? I can install Windows 7 if there are windows based GPU memory testing software.

0 Likes