Total and free memory Reported free memory differs between the driver and runtime API in 3.0

Accelerated Computing CUDA CUDA Programming and Performance
1

I am using the both the driver and runtime API calls to determine the available total and free memory of my GTX-260s. While the total memory reported agrees between both APIs (more or less), the free memory reported does not. I am using ctypes Python bindings. Has anyone seen anything similar? The output I get, is shown below. I initialize the runtime API and then create a temporary context to use the driver API call (as suggest once somewhere in a post by tmurray?).

[codebox]---- Memory reported by runtime API ----

Free [total] memory in MB = 513 [896]

Total memory: 939196416 = 896 MB

Free memory: 538390528 = 513 MB

---- Memory reported by driver API -----

Total memory: 939327488 = 896 MB

Free memory: 906100736 = 864 MB

[/codebox]

2

Uh I don’t think that’s actually what I suggested. There’s no reason to do that anymore (or ever, really, unless you were doing this at the beginning of an app). If you’re using Vista, then you’re hitting Vista paging behavior.