Graphics AMD’s DirectX 12 Advantage Explained – GCN Architecture More Friendly To Parallelism

things are moving so fast, that the comment you linked is now redundant - they can now run 10 wavefronts....

and the 970 does use segmented ram - NVidia said so themselves (on the final cuda block)

edit:

would seem that even AOTS isn't pushing the GCN 1.2

 

Now you've thoroughly lost me! You sure you're not getting GCN 1.0 mixed with GCN 1.1/1.2?

The 7970 has 2 ACE, they are both single threaded. Surely that means they can only handle 1 queue each? The quoted paragraph, and this is how i interpreted it, says

IMO the they is referring to the Southern Island Devices, not the ACEs

GCN 1.1/1.2 have, up to, 8 ACE each containing 8 queues. E.G. the 290x has 8 ACEs, each having 8 queues (threads?) for a total of 64 queues.

The other reason I think that GCN 1.0 only has 2 queues is, on that Anandtech article where everyone is up in arms about the GCN/Maxwell comparison, Zlatan didn't correct Ryan on GCN 1.0 he left that as 2 Compute.


Anyone else notice the AMD PDF said ACE engines? (Asynchronous Compute Engine engines )

 

Wavefronts?
Was that directed at loftie because I don't have a clue what a wavefront is, other than the weather variety.

Yes, but people were using it to not only to detect the segmented RAM but also to benchmark it, something the program wasn't designed to do, IIRC it bypassed the memory management in Nvidias DX11 drivers and directly addressed the card using CUDA (Nvidia's own proprietary parallel computing platform and API).

Kind of the same thing this new program does, all it does is generate Async compute in CUDA, it's a quick and dirty way to test if a card can do Async compute, not to measure how fast it is or anything else as it bypasses all optimisation work done via drivers (DX11)

 

then maybe you want to actually read the links you supply - wavefronts are mentioned later in that very thread....

still doesn't explain the stuttering people still get in games on 970`s that are not on 980`s...

and its not the same - AOTS is a game not something `quick and dirty` - but please , go over to B3D and tell them you have far more skill in creating programmes designed to test part of the DX12 feature set than they do , at short notice , which can produce verifiable results on multiple platforms.

AMD have created something which has potential - but why all the extremism? its not as If they suddenly made nv redundant , they simply have leveld the playing field for now , till nv find yet another way to cheat.

 

So your talking about the part that I initially ignored because it related to the strange results with a Fury X? I fail to see how that's relevant to the use of a program that's not designed as a benchmark being used exactly for that purpose.

Perhaps you care to point out, for us slow folks, what a wavefront is and why they're relevant to a program written to specifically test for the existence of async compute.

Since when have we been trying to explain why stuttering happens?
I though we were discussing that using a program not intended as a benchmark to benchmark something was an indication of absolutely nothing, other than IF the card can do async compute.

Wouldn't getting into a discussion on why some people supposedly get stuttering, something that's still very much debatable due to the crazily high setting used during tests, wouldn't that be taking this conversation off on a tangent?

Sorry I'm failing to see why you're getting so angry, the section you quoted from me makes no mention of AotS so IDK why you brought it up and used it as a stick to beat me with, I was under the impression we were discussing the validity of the program MDolenc wrote to test for the existence of async compute.

I'm also not sure how you're defining testing parts of DX12 on multiple platforms as the program has been re-written six times to deal with different architectures, there's plenty of doubt on the B3D forums as to the validity of using such a program as a benchmark if you cared to read about it.

The only extremism I see is in your reaction to what until now had seemed like a reasoned discussion on the differences in architecture between AMD & Nvidia, and what impact the different designs have on DX12 versus DX11.

 

Reading through some of the links posted Async is bad for VR with latency times in the 60-70ms instead of 15ms. If VR is as badly affected by latency as some think it will be you could end up with some interesting choices between VR or Async.

Most of the nvidia tests the card is sub 10ms indicative that it's not doing Async. AMD all cards are at least 50-60ms and that increases as the workload does. One of the devs for VR was posting in the topics saying anything above 20ms would effect VR.

That would be worrying for any would be VR user

 

If you're basing that on the charts/graphs than show GCN has a flat line, and Maxwell as a lower gradually increasing line, don't because iirc the guy who wrote the async test program stated it's not designed to test the latency, it's only there to test if async compute is working. Think the post was linked to on the Anand forums.

Personally, as AMD stated, no card really fully supports DX12 so I'm far more interested to see the next set of cards, but then I would be regardless

 

Thus far, the 'latency impact' of Nvidia architectures (listed variously as 60-70ms, 33ms, 24ms, etc) does not appear in practice. Fire up Oculus World and check out the latency debug overlay.
A lot of people are seeing "asynchronous compute shaders" and "asynchronous timewarp" and thinking the two are intimately related because they share a word. They key arbiters of Asynchronous Timewarp's effectiveness are the accuracy of Late Latching timing (timing before the next VSYNC readout begins) and making sure the GPU is not busy at that time (don't use draw calls that take super-long to complete*).

* Basically, if the GPU is taking 8ms to complete a draw call, there's a chance that the time to perform the Async warp comeas and goes while the GPU is still working. If you split that call into 2ms chunks, then every 2ms there's the chance to pre-empt the next chunk and fire off the Async warp. This is a less a "OMG, this architecture is no good for VR!" and more and "oh, that's good to know for optimisation among a whole load of other low-level considerations because VR performance optimisation is hard".

 

David Kanter talking about async compute, AotS, and AMD's ACE engines on a TR podcast.

If it's not jumping to relevant part they start talking about it at 1h 12m 30sec and then prattle on about HPC until 1h 19m when they actually start talking about AotS and async compute.

 

Asynchronous compute, AMD, Nvidia, and DX12: What we know so far
https://www.extremetech.com/extreme/213519-asynchronous-shading-amd-nvidia-and-dx12-what-we-know-so-far

Having read through this article is seem a fairly fair appraisal of the situation to me, nicely summed up by the following...

 

New Fable DX12 benchmark out

 

Doesn't seem UE4 makes extensive use of CPU threading, which is highly disappointing considering how oft-used UE is. Someone needs to contact UE team to get more insights into their DX12 implementation because it doesn't look like it's using everything available from the Anand benchmarks.

 

also it appears that the engine has certain features `switched on` for consoles but not for desktop ; which doesn't make sense tbh

 

CPU threading isn't an easy thing to use. Many tasks simply can't be threaded effectively. If you're spending 10ms on single-threaded engine simulation and 2ms on job dispatch, then cutting down that job dispatch time through threading by 4x only takes you from 12ms to 10.5ms of CPU time before render starts.
DX12 is not, and never will be, a magic go-faster switch.

 

The fable results are not a huge shock it was never draw cell limited on dx11. And the main benefit of dx12 is draw cells.

Rts games will be where people will see huge dx12 gains at least in early titles. Be mid next year before a PC game will come with most major dx12 features included be that long before there's a gpu with all the features in truth.

 

When is next Civ due? Should be soon - V was out in 2010.

 

and BE was last year