CUDA Memory allocation issue on GTX 850M on Windows 7 (64)

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi guys,

We have run into a weird issue where we are unable to allocate more than 2GB of memory on 850M that has 4 GBs. Just to make sure that there is nothing wrong with our program, we wrote a small test app that keeps alloc’ing memory until it receives out of memory error. It works fine on Windows 8, however on Windows 7, it throws error after allocating 2 Gbs, and yet it shows it has 2 Gb free memory available. My question is why I am unable to allocate more than 2Gb memory on Windows 7? Another clue: I rolled back my driver and found out that I get this behavior starting from nvidia driver 352.86 all the way to the latest driver. On older drivers, it works like a charm even on Win 7.

Thanks in advance.

Best

2

If the driver’s behavior has changed between two driver revisions, you might want to file a bug. You have to be a member of nVidia’s registered CUDA developer program ( there’s no cost involved, just a a registration ).

This may be related to how the Windows Display Driver Model (WDDM) manages memory. I just find it odd that there has been a change between specific CUDA driver revisions.

3

Based on your description, this sounds like a regression in the driver. I would suggest filing a bug report to bring this to NVIDIA’s attention.

The bug reporting form is linked from the CUDA developer website, log in at [url]https://developer.nvidia.com/[/url]

4

Edit: cudaMemGetInfo(free,total) on an sm_52 device doesn’t seem to adjust the free count after a single very large allocation. Yet the free count is adjusted on the two sm_50 devices in my 358.50 + Win7/x64 system.

If this is acceptable behavior then the function’s description should probably spell this out.

@Usmann, what is cudaMemGetInfo() reporting?

You could try running this test – it’s probably similar to yours: Allocate more than 4GB · GitHub

If you haven’t already, you should verify whether you’re building a 32-bit or 64-bit application.

Example on a 4GB GTX 980 (sm_52) + 358.50 + Win7/x64:

32-bit arch

$> malloc
GeForce GTX 980 (16)
free/total (MB):    3066 /    3072
cudaMalloc (MB):    2810
free/total (MB):    3066 /    3072 <--- THIS IS UNEXPECTED

64-bit arch

$> malloc
GeForce GTX 980 (16)
free/total (MB):    3935 /    4096
cudaMalloc (MB):    3679
free/total (MB):    3935 /    4096 <--- THIS IS UNEXPECTED

Example on a 2GB GTX 750 Ti (sm_50):

32-bit arch

$> malloc 2
GeForce GTX 750 Ti ( 5)
free/total (MB):    1651 /    2048
cudaMalloc (MB):    1395
free/total (MB):     255 /    2048 <--- OK

64-bit arch

$> malloc 2
GeForce GTX 750 Ti ( 5)
free/total (MB):    1650 /    2048
cudaMalloc (MB):    1394
free/total (MB):     255 /    2048 <--- OK

A Quadro K620 w/2GB (sm_50) shows results similar to the GTX 750 Ti.

5

@cbuchner1, njuffa: Thanks for the reply

@allanmac: This is what I am getting:
GeForce GTX 850M ( 5)
free/total (MB): 4011 / 4096
cudaMalloc (MB): 3755
free/total (MB): 4011 / 4096

There is a slight issue in the code. You are not checking for the return code of cudaMalloc call, in my case I am getting memory exception and therefore nothing was actually allocated. So the third line makes sense and is valid. I am quite sure that on your GTX 980, its the same, you are unable to allocate anything and hence the free/total memory stays the same. Please slightly adjust the code to see if you get error on cudaMalloc on GTX 980.

@all:

I have logged a bug on NVidia drivers

6

Oops… Fixed!

However no errors are being returned and cudaMemGetInfo() is now reporting the expected result every single time.

The only difference between today and yesterday is that I rebooted! :)

Thanks for filing a bug.