When I build my computer, I used the graphic card of the very old one instead of buying a new one.
So I have a very fast Intel cpu (14 nm) without integrated graphics along a very slow gpu (40 nm).
Even with games released 5 years ago, there’s a lack of graphic processing performance.
I’m getting near equal speed with the Microsoft Direct3D software render and the old graphic card. An OpenCl (opencl also allow to aggressively use cpu extensions) benchmark program run at least 20 times faster on the cpu.
How this is possible and would be a benefit
There were a great improvement since the introduction of sse : the latest version of avx will provide 512 bits floating point arithmetic.
With the introduction of thin graphics api like Vulkan or Direct3D 12. writing api drivers shouldn’t be a large amount of work as it was for OpenGl or Direct3D 11
The situation is there are many software renders, some of them (like the one of Microsoft) even allow to run Direct3D programs. But the one I’m looking for would make an intensive use of simd instructions at the same level of OpenCl. As I also want to use programs that use Direct3D, I think the best way for implementing this would be a virtual graphic card with a wddm 2.0 driver.
It would be a bonus if that software would be able to combine hardware rendering and this kind of software rendering at driver level. Though it might only be possible with low‑level api like Direct3D 12.
There’s a supplemental reason for the virtual gpu way : with Vulkan and Direct3D, the performance would simply be added to existing hardware ones (however this require to be done inside the driver for Direct3D 11 or OpenGl which represent my supported hardware).