Julio Jerez wrote:I would be so nice if Vulkan exposes all these devices.
also I was thinking, that if the system come with an integrated on ship GPU, even if eh graphics are not up to the level of a discrete GPU, for high performance computing it may actually beat the GPU because the integrated GPU has to shares memory with the CPU.
so if there is a way to map the buffer this may be fart better that DMA transfer to a GPU via PCI bus.
I am thing of stuff like rigid body solve and collision system stuff that only run on CPU.
If I am right and this speculation, them CPU may has no lost the fight versus GPU, integrated graphics may can be use as math coprocessor fro highly intensive operation both in memory and operations.
My wet dream is that something like Vulkan exposes the integrated GPU and that we can use it just and the interface for high performance computing, disregarding graphics very much like CUDA on NVidia, but far more flexible.
You can use multiple GPUs with Vulkan (Next Vulkan will have big improvents there, but iguess this means graphics things like avoiding to decompress images for transfers. For Compute usege there's all there already).
The problem is: Intel has Vulkan drivers for Linux, but for unknown reasons not for Windows.
Thera is only a beta driver Go to their forums and get angry
Physics on integrated GPUs always seemed perfect to me, although there are 2 issues:
CPUs with high core count don't have integrated GPUs.
iGPU causes heat and bandwidth so slows down cores.
But you made a good choice to prefer quad core, let me know when you have some experience with this.
I also guess the 480 is maybe as fast as a 1080 in compute (based on dated experience from Kepler vs. GCN).
Update on async compute: Unlike the small sucessful test i did it fails to be a win in practice.
(See https://community.amd.com/thread/209805)
So, seems we still have to wait a bit until we can utilize our GPUs 100%.
Modern APIs are a great improvement, but there's still work todo...