Since 2.73 version Blender 3D/CUDA/Cycles is using about 800MB of extra vRAM in any scene, even with the default cube. Supposedly this is related to new versions of CUDA that are used to compile the kernel in Blender/Cycles:
Blender 3D developers have commented that perhaps the use of vRAM will be reduced when the AMD/ATI Kernel Split arrives to CUDA:
But that may take a long time until it happens. Currently on cards with less than 2GB of vRAM, Cycles/CUDA can be practically unusable.
So from nVidia side: Can you do something to try to solve this problem?
Reading pretty much the entire thread at the Blender developer site it seems that newer versions of Blender added new features that require more memory, among other things. Since the problem is not caused by CUDA, I do not see how NVIDIA could address it (short of transparently compressing memory behind the scenes, which is difficult to pull off as history has shown).
The straightforward solution I see is not upgrading to the latest version of Blender, and sticking with the previous version that needed less memory. Alternatively, if you absolutely need features present only in the latest Blender version, upgrading your GPU seems indicated, models with 2 GB of memory are quite affordable now.
I am puzzled by your comment about a fix possibly requiring “a long time until it happens”. I was under the impression that Blender is a popular open source project with quite a few active developers. One of the often touted advantages of such projects is fast turnaround for any fixes needed.