Nvlink and AMD data fabric should improve gpu to gou communication. At least from what I read the latency for AMD CDF is less than that of PCI-E which is a bigger factor than bandwidth in this situation.
But low abstraction API's were needed for a long time. We have always been cpu bound when it comes to all graphic situations. Devs should not have been jumping through hoops just to get something to work. Not a matter of optimisation, just them being forced to work within a tiny envelope compared to a console for draw calls.
Is why I give the example of GTA3 on DX9. It was not that the game was badly optimised, the game engine was very optimised to be able to work on the 360 and ps3. it was DX9 that needed the cpu grunt because of the game being open world and the tiny draw call limit of DX9.
So DX10 would have improved things from its higher draw call limit. But even DX 11 holds back GTA V.
We should have had these api's years ago in some form.
But low abstraction API's were needed for a long time. We have always been cpu bound when it comes to all graphic situations. Devs should not have been jumping through hoops just to get something to work. Not a matter of optimisation, just them being forced to work within a tiny envelope compared to a console for draw calls.
Is why I give the example of GTA3 on DX9. It was not that the game was badly optimised, the game engine was very optimised to be able to work on the 360 and ps3. it was DX9 that needed the cpu grunt because of the game being open world and the tiny draw call limit of DX9.
So DX10 would have improved things from its higher draw call limit. But even DX 11 holds back GTA V.
We should have had these api's years ago in some form.