Are you using the latest version from git or a binary release? I see a commit message on Dec 28th that says "add back support for chunked memory allocation and texture cache to Kepler kernel. Slight speed-ups with -C 1 are seen." which implies he removed it at one point, possibly when upgrading to CUDA 5.5.
Was using the latest binary, the 12-18 release. I'll try compiling from git, thanks.
edit: Is there a guide somewhere to compiling on windows? I got all the components but I'm unsure what to do with them.