Re: wined3d/directx.c: Corrected Nvidia 8400 and 8500 reporting
Hi Seth, It is not a good idea to advertise a 8400 and a 8500 as a 8600GT. All Geforce8 GPUs have the same features (except, some have different purevideo capabilities). As mentioned before even CUDA is allowed on all GPUs but only if they have 256MB or more. Likely the app you are using either disallows this GPU based on the wrong number of video memory Wine reports or it just disallows poor GPUs because it is not worth the effort to use them for CUDA computations because they don't have enough computation power. The Geforce 8600GT is (depending on the configuration) upto 4x faster (if you just look at shaders + clock speeds), so it is really not a good idea to mark the 8400/8500 as a 8600GT. This could really cause issues for games which use the PCI id to select a performance profile at startup. It would be better to add a separate 8500GT entry with 256MB. I'm not sure if we want to merge it with the 8400 though since the 8500 has 256-1024 (depending on the model) and the 8400 has 128-512MB. A lot of modern games like to have around 256MB. I guess it is best to keep the 8400 tied to the 8300. Roderick On Sun, Jul 11, 2010 at 7:04 PM, Seth Shelnutt <shelnutt2(a)gmail.com> wrote:
Moves the Nvidia Geforce 8400 and 8500 to be reported as 8600GT as they have feature parity. This is needed for CUDA applications to support these two cards. They are currently reported as an 8300 which is not CUDA capable.
Thanks,
Seth Shelnutt
Roderick, You are right in that cuda is allowed on all GPU with 256mb of memory or more, the problem is that the 8300 doesn't support cuda. The 8300m does, but having the 84/8500 report as an 8300 means that if a program, folding(a)homein this case, checks the gpu reported against a list of cuda supported devices, then it will fail. Weither it is because of the video memory reporting or simply the string "Nvidia Geforce 8300 GS" I am not sure. Separating the 85/8400 into it's own id is valid and easy enough to do. I had tied them to the 8600GT because I thought it was simpler and created less overhead for just two GPU's that probably don't represent a large population. I think the 8400 should be tied to the 8500 or on it's own. If you put the 8400 back with the 8300 then someone might run into the same issue we are having now. Thanks, Seth Shelnutt On Mon, Jul 12, 2010 at 2:51 PM, Roderick Colenbrander < thunderbird2k(a)gmail.com> wrote:
Hi Seth,
It is not a good idea to advertise a 8400 and a 8500 as a 8600GT. All Geforce8 GPUs have the same features (except, some have different purevideo capabilities). As mentioned before even CUDA is allowed on all GPUs but only if they have 256MB or more.
Likely the app you are using either disallows this GPU based on the wrong number of video memory Wine reports or it just disallows poor GPUs because it is not worth the effort to use them for CUDA computations because they don't have enough computation power.
The Geforce 8600GT is (depending on the configuration) upto 4x faster (if you just look at shaders + clock speeds), so it is really not a good idea to mark the 8400/8500 as a 8600GT. This could really cause issues for games which use the PCI id to select a performance profile at startup.
It would be better to add a separate 8500GT entry with 256MB. I'm not sure if we want to merge it with the 8400 though since the 8500 has 256-1024 (depending on the model) and the 8400 has 128-512MB. A lot of modern games like to have around 256MB. I guess it is best to keep the 8400 tied to the 8300.
Roderick
On Sun, Jul 11, 2010 at 7:04 PM, Seth Shelnutt <shelnutt2(a)gmail.com> wrote:
Moves the Nvidia Geforce 8400 and 8500 to be reported as 8600GT as they have feature parity. This is needed for CUDA applications to support these two cards. They are currently reported as an 8300 which is not CUDA capable.
Thanks,
Seth Shelnutt
I'm fine with tying the 8400 and 8500 together if you assign it 128MB of memory for now. I think the desktop cards typically have 256MB or more but some boards (and I think mostly laptops) might have only 128MB. Though the same issue would remain for the 8100/8200/m8300 regarding cuda, but those cards aren't used that much (well some cheap laptops/desktops might use them) but using them for CUDA might really not make sense just of performance reasons. Roderick
Separating the 85/8400 into it's own id is valid and easy enough to do. I had tied them to the 8600GT because I thought it was simpler and created less overhead for just two GPU's that probably don't represent a large population. I think the 8400 should be tied to the 8500 or on it's own. If you put the 8400 back with the 8300 then someone might run into the same issue we are having now.
Thanks,
Seth Shelnutt
On Mon, Jul 12, 2010 at 2:51 PM, Roderick Colenbrander <thunderbird2k(a)gmail.com> wrote:
Hi Seth,
It is not a good idea to advertise a 8400 and a 8500 as a 8600GT. All Geforce8 GPUs have the same features (except, some have different purevideo capabilities). As mentioned before even CUDA is allowed on all GPUs but only if they have 256MB or more.
Likely the app you are using either disallows this GPU based on the wrong number of video memory Wine reports or it just disallows poor GPUs because it is not worth the effort to use them for CUDA computations because they don't have enough computation power.
The Geforce 8600GT is (depending on the configuration) upto 4x faster (if you just look at shaders + clock speeds), so it is really not a good idea to mark the 8400/8500 as a 8600GT. This could really cause issues for games which use the PCI id to select a performance profile at startup.
It would be better to add a separate 8500GT entry with 256MB. I'm not sure if we want to merge it with the 8400 though since the 8500 has 256-1024 (depending on the model) and the 8400 has 128-512MB. A lot of modern games like to have around 256MB. I guess it is best to keep the 8400 tied to the 8300.
Roderick
On Sun, Jul 11, 2010 at 7:04 PM, Seth Shelnutt <shelnutt2(a)gmail.com> wrote:
Moves the Nvidia Geforce 8400 and 8500 to be reported as 8600GT as they have feature parity. This is needed for CUDA applications to support these two cards. They are currently reported as an 8300 which is not CUDA capable.
Thanks,
Seth Shelnutt
participants (2)
-
Roderick Colenbrander -
Seth Shelnutt