![]() ![]() ![]() With the same number of ALUs and nearly the same bandwidth you could directly compare how the generations are doing under Async Compute and if a stronger Geometry-Frontend potentially limits the gains under Async Compute because there might be less idle-times if the Graphics-Pipeline is less of a bottleneck. My main reasoning is of course the 380(X) (Gen 3) which shows basically the same results like the RX 480 (Gen 4), pointing out to bandwidth limitations as the main reason for less gains under Async Compute.Ĭomputerbase had the perfect test case for this question but didn't made any comparisons in regards to Async Compute gains: However, it's likely that we will have DX12/Vulkan only applications where we won't have direct comparisons in contrast to DX11/OGL, even if many if not all features could be achieved through IHV-Extensions and some different implementations.īut developing applications full with IHV specific extensions are of course not the most attractive for developers, if there is a common standard and at a certain point it will become unreasonable to try the same things under DX11/OGL.Ĭlick to expand.In comparison to the 390 and Fury you definietly see cases where the benefits are smaller but how much of it is directly related to the improved architecture and not due the configuration or bandwitdh? To be fair, there might be DX11 extensions available through NVAPI and AGS to mimic ExecuteIndirect and for OpenGL there are also extensions for bindless resources. Going from DX12 to bindless and then using ExecuteIndirect shortened the time by 18% respectively DX12 alone took 21% more time to process on the GPU side. Nonetheless the interesting observation about this example is, that the performance always improves on boths sides, even if small. Where of course the opposite is to be expected for DX12 applications but it's a simple demo running on a power shared SoC. There might be a third with Rise of the Tombraider since there is a Async Compute option in the config but I didn't find results for it.įortunately with Gears of War 4 there will be an upcoming title which also allows Async Compute on/off comparisons.Ĭlick to expand.You will also see that the switch from DX11 to DX12 in this example dramatically improves GPU performance but not so much CPU times. ![]() I would be interested in cases where you could see smaller gains from Async Compute on the RX 480 but there are currently only two applications which allow comparisons between Async Compute on/off and are publicly tested. I agree, the variance under AotS is too high for precise arguments but unfortunately there isn't much choice.ġ0,6 % for the 380X Nitro, 12,1% with the RX 480 and 12,8% for Fury X. SM6 may also come to DX11.3/4 but the explicit resource managament, the command-list-generation, ExecuteIndirect and the new binding model are only available under DX12, as long I didn't get anything wrong: DICE and Confetti presented methods for triangle culling and a visibility buffer, one key foundation are the expanded capabilities coming with ExecuteIndirect. ![]()
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |