Question about NVidia cards and CUDA and NVEnc settings

[Dirkmaster]

I have a NVidia 970 graphics card, which according to it's spec has 1664 CUDA cores. Yet in the hardware Optimizations, the Maximum Max instances I can set is only 16. If I set any higher number, it just falls back to 16. The same is grue NVenc.

Under conversion settings, there are "number of used cores", but this seems to be CPU cores. How do I set it so the program doesn't use CPU cores, but just uses the CUDA cores

I would LOVE to know what the correct settings are for my Graphics card, so that I can get the fastest possible processing. I am unclear what exactly affects what, and the documentation I have found has been light on details.

thanx

Dirkmaster

[JJ]

As you noticed, that max cores setting is for CPU.
Currently there is no way to set or change the number of CUDA cores within any program, maximum available is usually used.

[Dirkmaster]

Thanx for the reply, JJ. So, any idea what the Max Instances does then?

[JJ]

I suppose that each conversion process is one instance of Cuda.
You can convert several files at once, if you select "single file output".
Converting to single target (DVD, BR) limits amount of instances.

[Brainiarc7]

I have a NVidia 970 graphics card, which according to it's spec has 1664 CUDA cores. Yet in the hardware Optimizations, the Maximum Max instances I can set is only 16. If I set any higher number, it just falls back to 16. The same is grue NVenc.

Under conversion settings, there are "number of used cores", but this seems to be CPU cores. How do I set it so the program doesn't use CPU cores, but just uses the CUDA cores

I would LOVE to know what the correct settings are for my Graphics card, so that I can get the fastest possible processing. I am unclear what exactly affects what, and the documentation I have found has been light on details.

thanx

Dirkmaster

Hello there,

NVENC, unlike the previous CUDA implementation, runs on a dedicated silicon IP block on your GPU, in a fashion similar to Intel QuickSync's and AMD's VCE (Video Coding Engine) blocks, in that they are indeopendent of your GPU's shaders, and are thus, much faster and more power efficient in comparison to previous GPGPU - based hardware accelerated encodes such as CUDA's NVCUVENC APIs and OpenCL's Lookahead implementations in MainConcept and similar SDKs.
With NVENC on consumer GPUs, you're limited to 2 encoder sessions and on Quadros, the number may vary depending on the SKU.
There's no manual way to override this limitation as it's hard-coded into the firmware of the GPU, and the API only exposes generic initialization contexts (via CUDA or DirectX), and not direct access or control to the hardware's resource utilization limits.

Regards,

Brainiarc7.

[Dirkmaster]

Alrighty then. Thank you very much, sir, for that enlightening reply. Are AMD cards able to throw more resources at a job, so that more conversions can happen simulanteously?

thanx again

DIrkmaster

[Brainiarc7]

With AMD, two use-cases can happen:

1. With older (VLIW-based) AMD GPUs such as the HD 5000 series and some APUs, they'd be detected as single OpenCL platforms on the latest VSO Video Converter alpha (Under Khronos' OpenCL platform).
2. If this software supports AMD VCE (on newer AMD GCN series GPUs and APUs), then you'd see one or two VCE devices with a fixed number of concurrent jobs depending on the developer mode used (Hybrid encodes allow for extra sessions whereas discrete VCE mode allows for an extra DEM (Desktop Encode Mode) useful for capturing video data from cameras and monitors) AND an extra OpenCL platform if you're on an AMD Enduro platform with the discrete GPU being a GCN part.

Regards,

Brainiarc7.