Note: in my code I use the BufferedImage....getData method to get an int for the backbuffer BufferedImage which I drawImage onto the BufferStrategy gc every loop, but I only get the int  once at the beginning then keep modifying the array. Just wanted to point that out from Abuse post. Probably a unmanaged, nonvram image is good in this case.
I've been trying these demos on various machines here at work (they're all good testcases for us, especially for the new OGL pipeline). I'm finding that performance for nonnus29's testcase is relatively poor with OGL enabled, but I think that's because we're using OGL to copy a software (unaccelerated, non-managed) image to the backbuffer and flipping on every frame.
Just to back up what you and Abuse suspected, managed images will no longer be accelerated once you call getRaster() or a related method. In this context, I don't see why you need to modify any image arrays directly. It would be great if we could see your source code. But from what I can tell, a more optimal approach would be something like:
- load bigimage.gif
- copy each tile from bigimage.gif into its own managed image (createCompatibleImage())
- render each tile directly into the BufferStrategy backbuffer (no need for an
- call strategy.show()
If you follow this approach, everything should be accelerated, and with OGL enabled, every tile will be cached in a texture, and the snowflakes will be alpha blended to the backbuffer all at hardware speeds. Let me know if this makes sense. It would be great to see an updated testcase.