Upcoming MAJOR memory optimization for CUDA GPUs

Run Cosmos Laundramat benchmark https://gooseberry.blender.org/gooseberry-production-benchmark-file/

If it fits into 6GB it will mean something…

What I read is that this feature is basically for use by developers.
But a few days ago has been another MAJOR CUDA memory optimization:
https://git.blender.org/gitweb/gitweb.cgi/blender.git/commit/999d5a67852b5958b9361c9888734ebc889e4a22

So if you use Blender from master (or Buildbot), you can get a huge reduction of vRAM. Tests to do would be:
Blender 2.77a
Blender from buildbot/master
Blender from buildbot/master + adaptative kernel

@dns_di, But that scene takes much more than 6GB of RAM with CPU. I do not see why it should fit in just 6GB vRAM.

My test with BMW27.blend scene and GTX 960:

Blender 2.77a: 1068 MiB
Blender from buildbot/master: 569 MiB
Blender from buildbot/master + adaptative kernel: 303 MiB

CPU
Blender open with the scene: 287 MiB RAM
Render: 530 MiB RAM