cudaHostAlloc not consistent between CUDA 3.0 beta 1 and CUDA 3.0

In CUDA 3.0 beta 1, I use cudaHostAlloc to allocate 700MB page-locked host memory and it works fine. However, in CUDA 3.0, the same allocation returns out of memory error.
I try to reduce memory size to 600MB, 500MB, 400MB it still return out of memory error. It just works when I set the size less than 300MB.
I understand the out of memory error, but thing I confuse is why cudaHostAlloc not work the same way in CUDA 3.0 beta 1 and CUDA 3.0. I wonder if CUDA 3.0 version sets a limit on page-locked host memory size that cudaHostAlloc can allocate.

Really appreciate anyone can help me verify the problem.