Alessio1989
Regular
I wonder now if the d3d12 final specification could have been a little different without Fermi...
Direct3D feature levels discussion
- Thread starter DmitryKo
- Start date
And regardless, I suspect Qualcomm is probably the limiting factor not Fermi.
That explains a lot.
Alessio1989
Regular
All the specification were rounded to match a well defined list of hardware architectures. This is why we have different capabilities tiers (and cap-bits) for different features and not only a couple or two of feature levels. Fermi was the oldest of the planned supported architectures ...
I am not sure nor aware about Qualcomm implementation, last time I asked there were no D3D12 drivers for W10M...
Mine is just a pure speculation..
https://www.qualcomm.com/news/onq/2...-microsoft-directx-12-and-qualcomm-snapdragon
Sure there isn't any driver support currently for WM10, but do you think Fermi is less capable than anything they have today (or even their next gen)? I suspect (granted it's pure conjecture) there are at least a few areas that qualcomm is the limiting party.
If I remember the FL 11_1 discussion correctly, Nvidia hardware had big limits regarding to UAVs (in graphics pipeline shaders). The 8 UAV limit (in Tier 1 resource binding) is likely because of Fermi. Qualcomm could of course have the same limit. Who knows, they aren't exactly open about thier architecture.
Alessio1989
Regular
Tier 1 of resource binding allows 64 UAV slots if the hardware support FL 11_1 (ie: haswell and broadwell). However all Tier1 GPUs in binding resources are not bindless (last time I tried on Haswell I had a runtime crash xD).
I am quite sure Fermi can bound only 128 srv to the pipeline (256 on later NV hardware), but I am not sure about other hardware, like intel Gen 7.5/8, ... or I am simply blending architecture limitations with CUDA? I do not remember how exactly CUDA references can be used to check real nvidia hardware limits.
Last edited:
Fermi is not a FL11_1 GPU if I remember correctly - it's 11_0.
But yeah I don't see the point in it being supported... probably just marketing that later got corrected by tech folks
But yeah I don't see the point in it being supported... probably just marketing that later got corrected by tech folks
Fermi is not a FL11_1 GPU if I remember correctly - it's 11_0.
But yeah I don't see the point in it being supported... probably just marketing that later got corrected by tech folks
A broader impression of incumbency and mind share due to the quantity of those cards out there?
There could be some cachet for Nvidia from a PR standpoint if "this old stuff *yawn*" could show up to the same fight as the competition's most modern architectures.
DmitryKo
Veteran
So Fermi is officially out now? That would explain its omission from this GDC presentation (see slide 42).
BTW Maxwell-2 is shown to have Resource Heap Tier 1 - but their driver was reporting Tier 2 in 2015 releases...
GDC 2016
Practical DirectX 12 - Programming Model and Hardware Capabilities
Gareth Thomas & Alex Dunn AMD & NVIDIA
http://32ipi028l5q82yhj72224m8j.wpe...ogramming_Model_and_Hardware_Capabilities.pdfPractical DirectX 12 - Programming Model and Hardware Capabilities
Gareth Thomas & Alex Dunn AMD & NVIDIA
We don't know this, mobile chips did have virtual memory support right from the start. Max McMullen said in this very thread that Microsoft was considering all mobile hardware vendors, not just Qualcomm, when making their specification, and WDDM 2.x requirements were known since at least WinHEC 2006.
It's just Qualcomm would rather invest their limited resources into Vulkan on Android, not WDDM 2.0 /Windows 10 Mobile which didn't take up.
Alessio1989
Regular
NV claims to have FL 12_0 GPUs on kepler which is false... https://developer.nvidia.com/dx12-dos-and-donts
Looks like they do not take care too much about d3d12 for now... lol I wouldn't never say that..
Fermi support was just marketing from the beginning... imo
Looks like they do not take care too much about d3d12 for now... lol I wouldn't never say that..
Fermi support was just marketing from the beginning... imo
Is there easy program for that? if so, I can
DmitryKo
Veteran
Either DXCapsView from the standalone Windows 10 SDK release 1511 (alternatively, an insider preview SDK build 14332)...
or, my console-based tool attached to this post: https://forum.beyond3d.com/posts/1840641/
Last edited:
Resource Heap Tier 1 is correct. Also mentioned here: http://on-demand.gputechconf.com/gtc/2016/video/S6815.html (21:30).
Tier 1 on these drivers (361.43)
Code:
Direct3D 12 feature checker (July 2015) by DmitryKo
https://forum.beyond3d.com/posts/1840641/
ADAPTER 0
"NVIDIA GeForce GTX 980"
VEN_10DE, DEV_13C0, SUBSYS_31701462, REV_A1
Dedicated video memory : 3221225472 bytes
Total video memory : 4294901760 bytes
Maximum feature level : D3D_FEATURE_LEVEL_12_1 (0xc100)
DoublePrecisionFloatShaderOps : 1
OutputMergerLogicOp : 1
MinPrecisionSupport : D3D12_SHADER_MIN_PRECISION_SUPPORT_NONE (0)
TiledResourcesTier : D3D12_TILED_RESOURCES_TIER_3 (3)
ResourceBindingTier : D3D12_RESOURCE_BINDING_TIER_2 (2)
PSSpecifiedStencilRefSupported : 0
TypedUAVLoadAdditionalFormats : 1
ROVsSupported : 1
ConservativeRasterizationTier : D3D12_CONSERVATIVE_RASTERIZATION_TIER_1 (1)
StandardSwizzle64KBSupported : 0
CrossNodeSharingTier : D3D12_CROSS_NODE_SHARING_TIER_NOT_SUPPORTED (0)
CrossAdapterRowMajorTextureSupported : 0
VPAndRTArrayIndexFromAnyShaderFeedingRasterizerSupportedWithoutGSEmulation : 0
ResourceHeapTier : D3D12_RESOURCE_HEAP_TIER_1 (1)
MaxGPUVirtualAddressBitsPerResource : 40
MaxGPUVirtualAddressBitsPerProcess : 40
Adapter Node 0: TileBasedRenderer: 0, UMA: 0, CacheCoherentUMA: 0
ADAPTER 1
"Microsoft Basic Render Driver"
VEN_1414, DEV_008C, SUBSYS_00000000, REV_00
Dedicated video memory : 0 bytes
Total video memory : 4294901760 bytes
Failed to create Direct3D 12 device
Error 887A0004: The specified device interface or feature level is not supported on this system.
FINISHED running on 2016-05-05 10:26:01
2 display adapters enumeratedSilent_Buddha
Legend
And yet the damage is already done. Because of NV PR most enthusiasts think it's level 12_0 and that Nvidia has better support for Dx12 than its competitors.
Regards,
SB
I have a feeling gamer people will fixate on the new hardware instead rather soon.
I am surprised Maxwell 1 isn't capable of more than 11_0. But hey it's not fast enough for future games anyway. Which is probably true of everything below GM204 and Hawaii. The common 2GB RAM allotment of recent times is also becoming a significant impediment.
By the way, Fermi does have WDDM 2.0 drivers. I have seen it myself with a 560 Ti and a GT 520. Maybe there are benefits to that compared to running WDDM 1.3 on Win10.
I am surprised Maxwell 1 isn't capable of more than 11_0. But hey it's not fast enough for future games anyway. Which is probably true of everything below GM204 and Hawaii. The common 2GB RAM allotment of recent times is also becoming a significant impediment.
By the way, Fermi does have WDDM 2.0 drivers. I have seen it myself with a 560 Ti and a GT 520. Maybe there are benefits to that compared to running WDDM 1.3 on Win10.
Last edited:
Kepler and Maxwell 1 are almost 11_1 capable. Only a couple minor features are missing. I fail to understand why Microsoft didn't exclude these features from the 11_1 spec. Because of this 11_1 is practically useless and we need to support 11_0 code paths for a long time.
2 GB memory will soon become the minimum. But you can't drop support of 2 GB cards, because top of the line Nvidia cards were 2 GB for a long time. Also Haswell IGP is limited to 2 GB of address space. 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.