http://www.overclock.net/t/1569897/various-ashes-of-the-singularity-dx12-benchmarks/1870






They did like I did at first. They looked at the documentation.
According to the documentation, Maxwell 2.0 supports Async
Compute/Shaders through HyperQ. It is very hard to find documentation on
Maxwell 2 because there is no White paper published by nVIDIA. This is
why I used the Kepler papers, wccf did the same.






The problem is that Beyond3D's test has shown that Maxwell 2 isn't using
Mixed mode at all. It is behaving like Maxwell instead. Maxwell doesn't
support Mixed mode. The Oxide developer stated that, as far as he knew,
Maxwell couldn't do Async.






The controversy has grown, rather than shrunk, with the tests Beyond3D
are doing because now their findings are point towards the ability of
Maxwell 2 to handle compute loads Asynchronously (32 Compute) but not
Graphics + Compute (31 Compute + 1 Graphics). This is what is most
interesting thus far. I don't think WCCFtech has grasped just what is
happening right now.






As far as a software scheduler, I also mentioned this in relation to
Kepler. One of the big changes between Kepler and Fermi was the removal
of a Hardware scheduler. This is why Kepler/Maxwell and Maxwell 2 use up
less power (not because they're designed more efficiently per-say). By
placing the scheduler in software, you can fine tune the scheduler using
the driver. This is why nVIDIA has far more leg room under DX11 to fine
tune the driver and derive a boost in performance. GCN, on the other
hand, relies on a Hardware scheduler. A Hardware Scheduler is better for
DX12, because the API is closer to metal leaving less room for shader
replacements and other forms of driver intervention.






What we have with Kepler/Maxwell/Maxwell 2 are cards which are fine
tuned for DX11. What we have with GCN are cards fine tuned for Vulcan,
Mantle and DX12.






Now while nVIDIAs Maxwell/2 cards can support more DX12 features (the
same was true of the GeForce 6800 Ultra) that doesn't necessarily
translate into better performance. The nVIDIA architectures lack the
Compute Parallelism performance now unlocked by the new APIs. For the
first DX12 titles, this might not be too much of a problem, but Pascal
would need to be a completely revamped architecture on this front. We
saw nVIDIA take the first steps with Maxwell/2 towards that direction.
AMD, on the other hand, are set to strike with an architecture which
will further boost their lead in this area.






We can't speculate as to whether Greenland or Pascal will be better, but
we can mention that AMD have far less architectural changes needed in
order to derive incredible Vulcan and DX12 performance. nVIDIA, on the
other hand, needs a huge overhaul of its architecture in order to
achieve the same result.