Direct3D feature levels discussion

sebbbi said:
7970 has aged really well. It now beats GTX 580 in most games, while it was certainly behind it at launch. 3 GB of memory + excellent compute performance is nice for current titles.
Click to expand...
Do you mean GTX680? 7970 was always way faster than GTX580.
 
Interested to see if GTX 1080 DX12 feature levels are equal to Maxwell Gen 2 or improved to say Intel Gen9 standards..
Hope some reviewer can post 1080 results here on 17 May a report using:
"Direct3D 12 feature checker (July 2015) by DmitryKo"
 
Interested to see if GTX 1080 DX12 feature levels are equal to Maxwell Gen 2 or improved to say Intel Gen9 standards..
Hope some reviewer can post 1080 results here on 17 May a report using:
"Direct3D 12 feature checker (July 2015) by DmitryKo"
 
Last edited:
I'm suprised that there is no GS bypass although Maxwell seems to support something similar already (In OGL).
FP16 is probably GP100 exclusive. (64 ALUs per SM, more shared-memory, FP16, FP64).
Personally I didn't expected RB T3 but I hoped for CR2-3 and Nvidia delivers CR2, so at least something.

Well now it's Polaris turn (Which I guess will be "dissapointing" too).
 
Infinisearch said:
Alessio why the hangup on the Resource Heap Tier? What so great about having different resource types on the same heap?
Click to expand...
Efficient aliasing techniques? Or just simply less overhead due different heaps indirections...

Locuza said:
I'm suprised that there is no GS bypass although Maxwell seems to support something similar already (In OGL).
FP16 is probably GP100 exclusive. (64 ALUs per SM, more shared-memory, FP16, FP64).
Personally I didn't expected RB T3 but I hoped for CR2-3 and Nvidia delivers CR2, so at least something.

Well now it's Polaris turn (Which I guess will be "dissapointing" too).
Click to expand...
As for RB-3, I hoped NV quitted with its sadomasochistic limits about constant buffer data and views :\
CR-2 is a nice improvement on CR-1, however Pascal looks as a good architecture in terms of performance to begin to see real serious stuffs with CR, unfortunately not everything is supported with CR-2 hardware (most important: SV_InnerCoverage semantic, which if I understood it correctly could give a nice performance improvement on fine grained occlusion techniques). But yes, CR-2 is a very nice improvement after all from CR-1.
 
Last edited:
Infinisearch said:
Is aliasing between a CB and a texture really a common thing to do?
Click to expand...
It is not only about buffers and surfaces, but also DS/RT surfaces... On tier 1 hardware of resource heap you cannot have different resource type in the same heap (https://msdn.microsoft.com/en-us/library/windows/desktop/dn986743.aspx). This limit your memory aliasing possibilities. I bet a complete general purpose aliasing opens the door of a lot of techniques and optimizations. Manual memory management is one of the key-features of D3D12 after all, and IHVs should not be afraid to give developers the possibility to implement such techniques just because they fear their driver will lose control of your hardware and so possible driver-side optimizations (the same reasoning apply to multi-GPU too, and at least one developer - Oxide - prove that developers can handle things better then drivers on complex software such today AAA games).

Infinisearch said:
I don't know what you mean by heap indirections, could you explain?
Click to expand...
I mean the need (of the driver) to dereference resource heap pointers (never try to manually dereference a gpu address!).
If you logically and carefully put together more resources in the same heap y (eg: access time, frequency access, aliasing pattern) you potentially decrease the number cache misses. Lowering the number of cache misses can be a real performance win.
 
Last edited:
AMD Polaris RX 480 (GCN Gen 4):

480_dx12.png


Drivers: 16.6.2
DirectX Caps Viewer 10.0.10586.122 (last non-beta Windows SDK)
Source: http://www.bitsandchips.it/9-hardware/7184-amd-radeon-rx-480-8gb?start=5

So no, nothings new here on D3D12 features.
 
One thing I was wondering - the 232mm^2 LinkedIn leak was real, what if the other one saying Greenland is first with GFX IP 9.0 was real, too?
That would mean Vega would more radically new, perhaps new D3D features and all?
 
You must log in or register to reply here.
Back
Top