What is most startling here is that Nvidia's own drivers tries to claim that Nvidia's GPUs has support for DirectX 12 asynchronous compute, which is something that Oxide says that their GPUs clearly do not really support. He also states the Nvidia only have a tier 2 DirectX 12 implementation whereas AMD has a tier 3 implementations, meaning that Nvidia has a little more CPU overhead.
http://www.overclock3d.net/articles/gpu_displays/oxide_developer_says_nvidia_was_pressuring_them_to_change_their_dx12_benchmark/1
To anand πάντα ήταν ενοχλητικά nvidia friendly (έτρεχε bench με πιο βαρύ antialiasing στις AMD από ότι στις nvidia και έγραφε "να δείτε, παίζει καλύτερα στην nvidia") αλλά δεν θεωρώ ότι λένε ψέμματα, απλά αναπαράγουν αμάσητο PR material της nvidia.
Γενικά μου φαίνεται παράξενο που τα άλλα 2 site που παρακολουθώ δεν έχουν καμία αναφορά στο θέμα (kitguru, hexus)
edit: διάβαζα στο forum του OC3D και έπεσα πάνω σε αυτό που εξηγεί τι παίζει:
<blockquote>Well, to be fair, Maxwell 2 does support Asynchronous compute. The Asynchronous Warp Schedulers do support this feature. The problem is in terms of context switching. When you enable Async Shading, on Maxwell 2, you have to limit the compute workloads. If you don't, you get a noticeable drop in performance. In essence, Maxwell 2 relies on slow context switching.
I don't know what's causing this issue. So I emailed Oxide, who came to our forums at Overclock.net, and explained the situation. This is where all of this info comes from.
I think it has to do with the L2 Cache shared by all of the CUDA SMMs. I think it might not have the available size and bandwidth to handle the load.
AMDs GCN, on the other hand, has a row of ACEs which reside outside of the Shader Engines. They can communicate trough the R/W L2 Cache, Global Data Share Cache and the GDDR5/HBM memory in order to synchronize, fetch and execute commands.
Another cause might be the DMA engines on Maxwell 2. They might not be as powerful or flexible as the ones found in GCN.
Just a theory.
In the end Maxwell 2 can support the feature, but loses performance when you work it too hard.
It reminds me of the GeForceFX. Which supported Pixel Shader model 2.0, but did so at a terrible performance premium.
</blockquote>