AMD: Speculation, Rumors, and Discussion (Archive)

Status
Not open for further replies.
Quite the contrary. It's very likely that host and shader code heavily optimized for GCN might heavily penalize NVIDIA HW and viceversa.


And this is something we have talked about many times over again and again and even have a dedicated thread for it, but still, the same stuff comes up every time a new game comes out or new patch comes out that shows async, its like that entire thread is totally forgot about, new slate.
 
From WCCFTech: "AMD RX 490 4K Gaming Card Listed By Sapphire & Spotted In Official Slide – Launching In 2016."

AMD-RX-490-Sapphire.jpg


WCCFTech speculates that the 490 is a dual Polaris 10.

I'm not saying it's not a dualie, but would they really name it RX490? That would be quite a departure from previous dual GPU naming schemes.
 
I'm not saying it's not a dualie, but would they really name it RX490? That would be quite a departure from previous dual GPU naming schemes.
The name has been very different for the past 3 generations, I don't think they have a naming scheme. 7990.. 295x2.. Radeon Pro Duo..
 
These complains are getting tiresome. Once it was "No game is using Mantle", than "no game is using DX12", than "No game is using Vulkan".
Mate, it goes back wayyyyyy longer than that... At least to the launch of the original Geforce, where it was "no game is using hardware T&L". Well - at least we're past THAT stage now! :LOL:

(Arguably you could say it goes back to at least the Matrox Mystique, where it was argued in whitepapers you didn't need things like texture filtering and alpha blending... ;))
 
One thing to always remember with ASC is that it kind of works like Hyperthreading, which means it helps you use idling processing cores in your chip. If you already run a high utilisation, there is not much to gain from it. If you could only use parts of your processing power before, it is a breakthrough. The Doombenches are still not much to celebrate if you consider that a Fury X has 8,9 GFlops while a 980ti has 5,6 GFlops.

So even there 59% more processing power only gives you less than 25% more FPS.
 
One thing to always remember with ASC is that it kind of works like Hyperthreading, which means it helps you use idling processing cores in your chip. If you already run a high utilisation, there is not much to gain from it. If you could only use parts of your processing power before, it is a breakthrough. The Doombenches are still not much to celebrate if you consider that a Fury X has 8,9 GFlops while a 980ti has 5,6 GFlops.

So even there 59% more processing power only gives you less than 25% more FPS.

Fury X is actually the worse example. Compare Hawaii and Polaris DOOM perf/flop to Pascals or Maxwells.

390X would be at 1070 level on that computerbase chart. That's 5.9 vs 6.5 Tflops.
 
Quite the contrary. It's very likely that host and shader code heavily optimized for GCN might heavily penalize NVIDIA HW and viceversa.
Not only that, but different iterations of the same or nearly the same architecture need to be balanced differently as well. So from what I've seen with AMDs nBody Simulation & AC, I'd say you have to carefully tune your compute threads for every chip supported in order not to run into a performance pitfall eventually. So, while I obviously welcome the added performance, I am also under the impression that this is not a fire-and-forget solution for the devs.

edit:
Coming to think of it, this is actually very similar to tessellation, where each architecture and sometimes each iteration of an architecture had their own sweet spots for performance, which sometimes shifted greatyl with drivres.
 
Last edited:
Instead of continuously turning this thread into a sh@tfest I think that we can all agree that, in your opinion, Async Compute, Shader Intrinsic etc...are a waste of time because they don't benefit NVidia. We all understood you very clearly I think... now let's go back on track and discuss AMD's tech which is the main subject of this thread.

The lack of perspective is notorious.

If we were complaining about tesselation or excess geometry being underutilized back in 2011, I wonder what the argument would be.
Main difference being that nvidia hardwires sub-pixel triangles into hairworks and other effects in gameworks, which serves only to cripple performance on AMD hardware and their own kepler/fermi cards, so we're not even talking about IQ-enhancing features anymore.

Now we have async compute that brings unprecedented performance boosts to the overwhelming majority of the console+PC market, has been acclaimed as such by pretty much all high profile developers (except for Epic because Tim and Jen are bros4life?), it's bringing major performance boosts to 4 year-old cards, but it's somehow a bad thing.
 
Love the ASUS info sheet in front of their RX480 with the wrong values against the fields for 'Base Clock' and 'Boost Clock', also the use of the term CUDA cores :)

EDIT: What's a semi-passive fan?

lol, look like the Asus Korean marketing team have got some problem to traduct the Taiwanese one.

Nice find, you was try to check the core speed ?
 
Here is the conundrum:

Q1:What is the purpose if these low level APIs?
A1:Increase the extracted performance of hardware, because abstracted APIs waste a lot of resources.

Q2: How is it then that these APIs cause negative performance on certain vendor's hardware, compared to the old "inefficient" API? Without even a single visual fidelity increase?
A2: !!!

It's quite obvious what we have here, either developers are dumb, NVIDIA drivers are so goos that they optimize close to the metal for every game on every GPU (both unlikely). Or these APIs are (at the current moment) simply a Mantle in disguise, designed to extract performance on certain vendor's hardware (AMD), while ignoring other vendors (NV, Intel.. Etc)

Q3: Will this help extend the lives of these new APIs?
A3: No, On PC they require extra time and work to be operational, and when you are targeting a limited set of hardware with the performance uplift, while causing havoc and fps degradation in the rest of the market for no logical reason whatsoever, then the majority of developers would rather stick to the old, relatively effortless APIs to do their games. Other vendors will be hesitant to join your initiative as well.
 
Status
Not open for further replies.
Back
Top