以下純討論...

剛才仔細看了該討論串，感覺還是有點迷惘，Beyond3D那邊有人做了一個
D12 Async Compute(以下簡稱AC)的測試程式，以證明NV的卡是可以支援AC
的，但有人提出反駁，認為樓主的資訊是有相當錯誤的，目前的情況是，有人認
為該程式由業餘人士所撰寫，而且是他第一個D12測試程式，應該參考長期專
業具知名度的程式設計者之意見才是正確的，其論點如下


People who read this should go to the SOURCE at beyond3D and read through to the rest for a PROPER interpretation because OP's interpretation is very wrong.
Async compute functionality would result in NO difference in time taken to complete the combined tasks of graphics + compute, because its running in parallel. What the programmers on b3d found, was Maxwell in fact cannot perform Async Compute, if you try to get it to, it reverts back to serial operation, graphics then compute, with the total time to completion the sum of those two tasks individually.
Really, go and read through, the programmers discuss what's going on and interpret the results much better than us laymen.


以下是資深程式設計師Sebbi的說明
https://forum.beyond3d.com/threads/...ad.57188/page-9
This is what Sebbi (very well known long time programmer on b3d) had to say about this program:

The latency doesn't matter if you are using GPU compute (including async) for rendering. You should not copy the results back to CPU or wait for the GPU on CPU side. Discrete GPUs are far away from the CPU. You should not expect to see low latency. Discrete GPUs are not good for tightly interleaved mixed CPU->GPU->CPU work.
To see realistic results, you should benchmark async compute in rendering tasks. For example render a shadow map while you run a tiled lighting compute shader concurrently (for the previous frame). Output the result to display instread of waiting compute to finish on CPU. For result timing, use GPU timestamps, do not use a CPU timer. CPU side timing of GPU results in lots of noise and even false results because of driver related buffering.
________________________________________

TLDR: This program only test for functionality, if Async Compute functions, the time to completion is shorter than the sum of doing compute and graphics individually, because its running at the same time. If Async Compute does NOT function, the time to completion is identical to the sum of the compute & graphics workloads done separately. Maxwell shows the latter pattern, it cannot do Async Compute.because its running at the same time.

小弟注意到這段話
If Async Compute does NOT function, the time to completion is identical to the sum of the compute & graphics workloads done separately. Maxwell shows the latter pattern, it cannot do Async Compute.
Edit: I've gone ahead and asked at B3D for more clarity, wait and see.

Beyond3D的討論串
https://forum.beyond3d.com/threads/...e-thread.57188/



如果這樣看來，Maxwell又似乎是不支援AC了(或說支援層級不如AMD)，AC算是D12
一個相當重要的特色，不知有沒有比較專業的高手能釋疑一下?