PS4's GPU is limited by it's memory bus and PS4 rendering resolution/AA. What's good for a PS4 isn't necessarily good for a PC GPU which may be limited by shading more than raster operations. As I've said the exact impact of that feature on GPU utilization on PC is unknown at the moment.
Now this is kind of crazy talk, the underlying GPU technology is the same, so the principial benefits.
This chart from anandtech is misleading, especially when looking at the context.
GCN Gen 1 / Southern Islands / IP v6 has two ACEs, each can dispatch one queue.
GCN Gen 2 / Sea Islands / IP v7 can dispatch up to 8 queues per ACE.
http://developer.amd.com/wordpress/media/2013/07/AMD_Sea_Islands_Instruction_Set_Architecture.pdf
Page 13:
Important differences between S.I. and C.I. GPUs
Multi queue compute
Lets multiple user-level queues of compute workloads be bound to the device
and processed simultaneous. Hardware supports up to eight compute
pipelines with up to eight queues bound to each pipeline.
Bonaire can handle up to 16 queues, in contrast to the first GCN Gen 1 product stack with 77xx/78xx/79xx which is limited to 2.
So writing simply two compute in the table is simply insufficient.
(The table was first claiming 1 + 1 and 1 + 7 in mixed mode for GCN)
GCN 1.0. I thought I've explained this.
I don't see where but I more likely believe that Fiji will have the same IP Version as Tonga.
GCN Gen 3 in the end.
GK104 was the first NV's GPU which can handle two compute queues. GK110 upped the number to 32 - four times more than 290 series 1,5 years earlier.
Any source about the cabilitity of the GK104 chip to handle two compute-queues?
Again, on PS4. PS4 is a non-conventional system in PC metrics. It have a lot of compute resources while having a comparatively narrow memory access path. If a GPU is starved by memory access in graphics queues then launching a pure compute queues in parallel can provide a substantial benefit.
I see nothing really special about it.
GCN Gen 2 put on a shared-memory bus, with access to 176 GB/s.
The CPU takes its slice, than you have some problems with memory contention and the effective bandwith for the PS4 GPU is under 140 GB/s, but from raw power perspective the 7850/7870 are in the same ballpark and featuring 154 GB/s.
The GPU doesn't need to be starved by memory access, bubbles in the work-queue and dependencies can always occur.
For GCN it should be principial needed, to use async compute to utilize most of the ALUs.
And with bigger GPUs I only see more bubbles and a greater need for async compute.
There's a bit of terminology mismatch since "queues" and "threads" are used rather interchangeably by different companies. I do believe that what AMD calls ACE is a queue processor in NV's terminology. So 8 ACEs in GCN1/2 and 32 of these in Maxwell 2 is what we have right now.
You can call it a queue-processor who can dispatch up to 8 queues.
Having 64 different compute-queues in total, when you build 8 ACEs.
If I look at hyper-q then Nvidia can handle 32 compute-queues with GK110.
http://www.pcgameshardware.de/screenshots/1280x1024/2012/05/gtc_2012_gk110_architecture_details_1.jpg
And if Anandtech got the right Info from Nvidia, Maxwell Gen 2 offers the same 32 queues.
