QUOTE(Demonic Wrath @ Jul 17 2016, 09:57 AM)
The Queue and Work Distributor is in hardware. The purpose of Work Distributor is to, as its name indicate, distribute work between SMs. If the SM needs to send its results back to the CPU to redistribute work between SMs, the latency cost would be too high. On a graphic pipeline, the data and instructions will need to go through the Work Distributor and the SMs multiple times. They wouldn't be able to get this kind of performance if they do this.
The driver queue, however, is in software. This is the same for both AMD and NVIDIA. The driver queue will send its data to the hardware queue to process.
Even the simplified look and the beyond3d post (by Ext3h) you linked show different things. Ext3h post indicates the queues, work distributor is in hardware. The simplified look indicates it is in software.
I would suggest you to go through the sources you link first before posting here to avoid misleading people..

dude even tech sites like pcper is for PR only, people from beyond3D are devs that program and test async on GPUs. i'm not saying they are entirely correct, nor that you are to take them as a fact, and anyone should not rely on what people say on the internet lol. did you read in the disclaimer that they also mentioned that it cannot be 100% correct, because all these info is gathered from limited testing/white papers/third party sources only, unless you know more than Nvidia engineers?
i just hope you stop misleading people with your own understanding, and say that Maxwell v2 Async works just the same as ACEs, because apparently it is not the same thing.
QUOTE(Demonic Wrath @ Jul 14 2016, 07:47 AM)
On NVIDIA GPUs, there's a hardware scheduler. It is the Gigathread Engine. This Gigathread Engine block is equivalent to (graphic processor + ACEs) in AMD GPUs. NVIDIA doesn't show the internals of their Gigathread Engine. In Vulkan, they have exposed 16 graphic queues and 2 copy engine in the latest driver.
This post has been edited by svfn: Jul 17 2016, 03:17 PM