AMD-The Master Plan

[R-T-B]

As I said, none of NVIDIA's cards are fully compliant with DX11. They accept the commands but they don't execute them how they were intended to be executed by Microsoft (asynchronously). AMD has for sure done it properly since HD 7###.

Ah you mean hardware compliance rather than driver level, gotcha.

In hindsight, it was kind of silly for me to assume otherwise really. If we accept driver level compliance the W10 generic VGA driver is actually the most compliant driver in town! (it implements all of DX11/12 in software render mode)

I wonder if the 10 series will fix the async issue in hardware once and for all...

 

[FordGT90Concept]

It better. If it doesn't, I'm going to laugh, and laugh, and laugh some more.

 

[TheGuruStud]

I'm sorry, but how does "The Master Plan" to defeat NVIDIA, not involve NVIDIA?

This was doomed from the getgo.

Also, x86 isn't and actually until recently wasn't the only game in town for consoles.

Bingo. It should be a Final Solution, not some half-baked plan.

 

[RejZoR]

Yeah, consoles had Cell and that IBM's tri-core thingie. And they were both too complicated for game devs. They had to specifically re-code entire games to even release them on those consoles. And doing the same the other way required same investment of time, money and manpower. And only ones who actually made good use of it were game devs who had Sony as publisher and who were involved in development of several titles for that console before thy got hang of it. Today, with unified architecture which is essentially PC crammed into a console case, things are a lot easier. They aren't any better for PC other than the fact we get console ports faster to and from consoles, but that's better than never getting them both ways because it was too much work and in the end it still ran like crap when they did it.

 

[RustyKats]

On this topic, another interesting video. I think it's from a developer, not sure but it gets technical and covers why Pascal is more future-proof than Maxwell.

 

[RejZoR]

What I'm more interested in is if NVIDIA has fixed the async shaders. Because buying GTX 1080 just to realize it still has crippled async compute, that would be just silly thing. Performance or not.

 

[RustyKats]

What I'm more interested in is if NVIDIA has fixed the async shaders. Because buying GTX 1080 just to realize it still has crippled async compute, that would be just silly thing. Performance or not.

You need a hardware scheduler and separate engines for that I think. Pascal's whitepaper does not refer to these inclusion at all so I doubt they have it.

But the above video goes into detail why Pascal is better at graphics and compute workloads, despite lacking "async shaders".

Personally, async shaders is a 10-30% performance gain potential in the few games that use it. Now, it's not a huge deal, this can be made up for with the fact that Pascal clocks a lot higher. Custom boards are talking about 2.5ghz overclocking.

 

[FordGT90Concept]

Pascal has the Gigathread Engine which should theoretically manage async tasks. Was it at fault of Fermi and sons failing at async shading? I don't know. Pascal is no doubt a step in the right direction but whether or not that step is big enough remains to be seen.

Async shaders are a paradigm shift from sync shaders. It isn't just about performance: it's about scalability, multitasking, and efficiency.

 

[RustyKats]

I think the Gigathread Engine is the same since Kepler. Fermi had a hardware scheduler, which NV removed for Kepler onwards to save on power usage (Fermi was a power hog).

Actually if you read Anandtech's Kepler review, they go into that in more detail.

 

[Caring1]

I thought I had read the new cards were capable of async, just that it hadn't been implemented.

 

[FordGT90Concept]

Maxwell accepts async commands because the DX11 API requires it. It doesn't actually execute it asynchronously. If the GPU is working on graphics and you throw a compute task at it, the whole damn compute unit has to shut down to change tasks. GCN, on the other hand, will happily do both if there's some hardware available to do it.

I think NVIDIA's goal was to make everyone buy a separate card for compute tasks so the graphics pipeline is never interrupted. NVIDIA apparently doesn't know what "async" means. They do now, at least.

 

[medi01]

Fact is anyway, these consoles should seize to exist. They are a pest for the PC users. Why own a limited environment PC like the XBOne/PS4 if you can have a proper one for the same money?

There are many, but I think one reason alone is more than enough: because on consoles you can sell/buy used games.

 

[RustyKats]

There are many, but I think one reason alone is more than enough: because on consoles you can sell/buy used games.

And you can't actually have the same pc for the same price. You need at least a 960 to match the gaming performance of the consoles due to the extra overhead on PC.

 

[RejZoR]

There are many, but I think one reason alone is more than enough: because on consoles you can sell/buy used games.

Can you really? Considering all consoles these days have login accounts and I'm pretty confident in saying a lot of games can't be traded anymore once they are tied to your account. I don't own any of the modern consoles, but seeing the state of greed companies are in, it wouldn't surprise me.

 

[medi01]

I'm pretty confident in saying a lot of games can't be traded anymore once they are tied to your account.

Multi-player focused - maybe. (even there I'm not sure).
Single player - no way.

 

[Kanan]

There are many, but I think one reason alone is more than enough: because on consoles you can sell/buy used games.

What are the "many's"? I think whatever reason you will tell me, will not be good enough for me to accept consoles that are solely made to earn more money than with PCs (through a different system, forces you to buy stuff at higher prices compared to PC - it's like having a Apple, just a tad better). Thats the real reason. You can exchange games on PC too - a PC can do ANYTHING a console can, and more. That's one of the most obvious differences of PC and console.

 

[FordGT90Concept]

Consoles are the lowest common denominator which is why they are usually the focus of development. Multi-platform titles usually sell more games on console than they do on computers. Console titles, before they can be published, have to be validated by the console manufacturers where there is no such oversight on computer launches.

AMD dominates the console market now. Making the code work on PC is damn near effortless--as long as the PC has a GCN GPU.

 

[FordGT90Concept]

PlayStation 4.5 could easily have dual GPUs. One GPU is disabled when running in legacy mode while 4K titles use both GPUs. When PlayStation VR is in use, each GPU renders an eye.

PlayStation 4.5 is scheduled to launch before PlayStation VR which launches in October.

Source: http://www.cnet.com/news/new-ps4-what-we-know-about-sonys-rumored-ps4-5-playstation-upgrade-ps4k/

The drop to 14nm makes a dual GPU solution affordable to AMD and Sony. This also explains why AMD is pushing out a mid-range chip (Polaris) instead of a high-end chip (Vega) like NVIDIA is (Pascal). AMD already has bulk orders for dual-Polaris PS4.5 APUs and presumably single-Polaris APUs for Nintendo NX. Sucks for PC gamers but this pretty massive for console gamers, VR, and AMD.

Sony is undeniably going to leverage Vulkan and LiquadVR to suck every drop of performance out of the dual-Polaris configuration.

 

[Kanan]

PlayStation 4.5 could easily have dual GPUs. One GPU is disabled when running in legacy mode while 4K titles use both GPUs. When PlayStation VR is in use, each GPU renders an eye.

PlayStation 4.5 is scheduled to launch before PlayStation VR which launches in October.

Source: http://www.cnet.com/news/new-ps4-what-we-know-about-sonys-rumored-ps4-5-playstation-upgrade-ps4k/

The drop to 14nm makes a dual GPU solution affordable to AMD and Sony. This also explains why AMD is pushing out a mid-range chip (Polaris) instead of a high-end chip (Vega) like NVIDIA is (Pascal). AMD already has bulk orders for dual-Polaris PS4.5 APUs and presumably single-Polaris APUs for Nintendo NX. Sucks for PC gamers but this pretty massive for console gamers, VR, and AMD.

Sony is undeniably going to leverage Vulkan and LiquadVR to suck ever drop of performance out of the dual-Polaris configuration.

If that is true, I would say it doesn't suck for us, it's instead good for us, because AMD needs some $$ and we profit from a better competition indirectly. Even if not, I would still like it, they deserve better times.

 

[TheGuruStud]

Dual would just complicate production. It's a lot easier to just beef it up by 50%.

It's woefully underpowered for VR, regardless

 

[FordGT90Concept]

Actually, it makes production easier/cheaper because yields are higher. There's less waste silicon. It makes the game engines more complicated though. I think Vulkan might be designed for dual GPU more so than DirectX which is why Sony will be the first out the door to utilize it and why the same tech won't be coming to discreet cards until 2018. 2018 is likely about the time PS4 titles will leverage the dual-Polaris design which will be ported to Linux, OS X, and Windows using Vulkan which, I don't think it is mere coincidence, is when Navi chips are supposed to debut. Navi could have 2, 3, 4, maybe even more GPUs on one interposer.

Dual Polaris GPUs should be plenty to run PlayStation VR. I would say it's overkill but they'll run it at lower clocks so it's really a perfect fit for PlayStation VR + console TDP.

It sucks for us because NVIDIA is going to be unopposed in the high-end graphics card segment until October at the earliest (Vega launch).


If all of this speculation is accurate, GPU tech is going to get very interesting over the next few years.

 

[32257870]

AMD controls the console market and DirectX12 so most if not all games that will come out in the future will be optimized for AMD cards/architecture and nVidia will slowly die.

The one thing that seems to keep AMD ahead of the game, even if things start to look bleak. AMD will have to screw up big time to NOT squeeze out Nvidia from the market.

 

[RustyKats]

For those interested about Polaris, a nice video that explains some of the features:

 

[FordGT90Concept]

I don't buy it. Polaris 10 will probably be around Fury X in performance which is a long ways behind GTX 1070. AMD knows this which is why hot on the heels of GTX 1070 being announced, AMD announced they're moving production of Vega (4096 stream processors) forward. In terms of hardware, Polaris 10 either matches or is slightly behind 390X; it makes up the difference with higher clock speeds.

And he's wrong: DirectX 11 DOES recognize the ACEs. Async compute shaders were part of the D3D11 specification which GCN 1.0 is fully compatible with. The problem was that APIs prior to Mantle were single-threaded so programmers couldn't feed those ACEs because the CPU thread to the API is tied up doing graphics tasks. DirectX 12 changes command queues to command lists which gives the GPU a list of tasks to gnaw on asynchronously:
https://msdn.microsoft.com/en-us/library/windows/desktop/dn899217(v=vs.85).aspx

 

[RustyKats]

I don't buy it. Polaris 10 will probably be around Fury X in performance which is a long ways behind GTX 1070. AMD knows this which is why hot on the heels of GTX 1070 being announced, AMD announced they're moving production of Vega (4096 stream processors) forward. In terms of hardware, Polaris 10 either matches or is slightly behind 390X; it makes up the difference with higher clock speeds.

And he's wrong: DirectX 11 DOES recognize the ACEs. Async shaders were part of the D3D11 specification which GCN 1.0 is fully compatible with. Game developers simply never used them until now because NVIDIA didn't put hardware support behind the concept.

Er, every dev talk on DX12 and such mention no such thing as DX11 capable of Async Shaders.

In fact, the terminology did not even exist until Mantle showed up.

The question is how fast do you think the 1070 will be? Much faster than Titan X? Because Fury X is ~= Titan X at 4K btw, and only a little behind at 1440p.