Async Compute disabled for GCN1 in current drivers

[Vesku]

At least in DX12:

While async compute worked fine in *almost* all cases on GCN1, we identified some corner cases where it could cause severe issues. We’ve discussed several ideas how the driver could identify and work around those corner cases, but for now the safest solution is to disable async compute on the affected cards. Sorry.

http://gpuopen.com/concurrent-execution-asynchronous-queues/#comment-6937

 

[ThatBuzzkiller]

So it's a hardware design bug ?

No worries, shader model 6 is more important than async compute for them ...

 

[Bacon1]

It is still enabled in Doom / Vulkan. I don't recall any games that saw a benefit with it for GCN 1 anyway.

http://www.anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta/6

2% diff, which is within margin of error as we saw fury x take a 2% perf hit with it on at that resolution.

This one from computerbase showed it having some gains, but not very playable settings:

https://www.computerbase.de/2016-02/ashes-of-the-singularity-directx-12-amd-nvidia/3/

Has anyone tested AOTS to see if it is actually disabled in the game or just a driver default of disabled?

 

[Alessio1989]

So it's a hardware design bug ?

No worries, shader model 6 is more important than async compute for them ...

I am not sure if GCN1 will support SM6. It is only mandatory for FL12_0+ GPUs. Moreover the current status of the driver causes synchronization and stability issues if you do not use an ad-hoc code path, like it was 1 year ago for Maxwell GPUs (now commands serialization under those NV GPUs works fine).

 

[Red Hawk]

Really now. That's unfortunate, but it makes sense to me. I've tested async compute in 3DMark Time Spy and Gears of War 4 on a 270X, and there was no performance improvement with the former and actually a slight performance loss with the latter.

 

[Yakk]

My understanding is GCN1 GPUs like Tahiti cards have 2 ACE units so while they are indeed Async compatible, they usually see little-to-no performance gain as both ACE units are usually always in use. The big thing is they don't have a performance loss when Async commands are sent to the GPU as they are just treated in sequence with, very importantly, no context switching required. So Async on or off should have the same results, with possibly a minimal gain on a badly coded path.

Can't see how this won't be fixed as all the hardware required is there.