2016 SIGGRAPH Khronos Videos To Watch About OpenGL, Vulkan, OpenCL, glTF

[phoronix]

Phoronix: 2016 SIGGRAPH Khronos Videos To Watch About OpenGL, Vulkan, OpenCL, glTF

The Khronos Group has now uploaded all of their session videos from last week's SIGGRAPH 2016 conference to YouTube...

http://www.phoronix.com/scan.php?page=news_item&px=2016-Khronos-SIGGRAPH-Videos

[Marc Driftmeyer]

Lots of useful information, but I thought I'd stir the pot with this point of the presentation from id Software:

https://youtu.be/CsHMiEQgrLA?t=4038

[chuckula]

Lots of useful information, but I thought I'd stir the pot with this point of the presentation from id Software:

https://youtu.be/CsHMiEQgrLA?t=4038

Eh, the first real DX12 benchmarks that implement asynchronous compute properly have shown that modern Nvidia hardware supports it just fine. The relative performance gains just aren't as great because NVidia has an inherently better scheduling system that doesn't leave unused computing resources lying around so there are fewer gaps in the pipeline to fill in with manual asynchronous compute programming.

[juno]

Eh, the first real DX12 benchmarks that implement asynchronous compute properly have shown that modern Nvidia hardware supports it just fine. The relative performance gains just aren't as great because NVidia has an inherently better scheduling system that doesn't leave unused computing resources lying around so there are fewer gaps in the pipeline to fill in with manual asynchronous compute programming.

What point are you trying to make?
D3D12 is not Vulkan. Nvidia has not yet enabled AC in their Vulkan driver, apparently.

[Marc Driftmeyer]

What id Software was inferring is that no matter what Nvidia does their will be latency overhead because they do not have the native hardware architecture to leverage full Async Compute on the level that AMD has built-in. With more and more resources leveraging Async id Software is stating flatly that the performance on AMD will only increase.

[juno]

What id Software was inferring is that no matter what Nvidia does their will be latency overhead because they do not have the native hardware architecture to leverage full Async Compute on the level that AMD has built-in. With more and more resources leveraging Async id Software is stating flatly that the performance on AMD will only increase.

Axel did not say that.
What he said was, that they are not sure, if they could present the finished image from out of the compute queue. The are not sure, b/c NV has not enabled AC yet in the driver. He continued that, if this was not possible, you'd need to sync the finalisation of the image, which is done inside the compute queue, w/ the graphics queue and then present from there. It is obvious that this switch is adding latency compared with presenting directly from the compute queue, isn't it? Still, what chuckula said doesn't have to do much with this.