Is there an in depth article comparing the AMD GCN arch and Nvidia Maxwell arch

[William Gaatjes]

I am looking for an article in the style of writing from David Kanter from Realworld tech.

I just find it strange that the fury has such a wide memory bus dividing the memory in sections for the different computing units. This should give it massive parallel processing capabilities. I am just wondering about how the fury has all the memory bandwidth it needs but it seems it cannot use that bandwidth effectively.

I am interesting in a good article that goes in depth about the architecture.
Anybody got some good links ?

Should i be looking in the beyond 3d forum ?

 

[Sequences]

I don't think you can find anything that parallels David Kanter's analysis except on his website, unfortunately. That's why someone offered him a job.

 

[William Gaatjes]

Yeah, i really like his articles. Even when it is difficult to grasp (sometimes i have to constantly stop reading and search for subjects he is mentioning to understand what he is talking about ).

But i found this as, it is a start . I will just add links to this post as i find them.
http://www.guru3d.com/articles-pages/amd-radeon-r9-fury-x-review,7.html

That way i and other members interested will have something to read and it is a central gathering point.

EDIT : This is interesting :

For the compute side of things, the GCN Compute unit model was introduced, it is designed for better utilization, high throughput and multi tasking. E.g. performance, performance, performance. Each CU, or building block, for Fiji is identical to what you have seen with Tonga.

 

[Silverforce11]

From reading discussion by gamedevs on this matter, GCN is running crippled on DX11 due to its serial nature, and it needs a very parallel API (Mantle, DX12, Vulkan etc) for it so shine.

We shall see soon enough with the wave of DX12 games incoming.

 

[monstercameron]

From reading discussion by gamedevs on this matter, GCN is running crippled on DX11 due to its serial nature, and it needs a very parallel API (Mantle, DX12, Vulkan etc) for it so shine.

We shall see soon enough with the wave of DX12 games incoming.

Allegedly the ace units are practically dead silicon running dx11 hardware. That's why Dan Baker of oxide fame has been tweeting about "free" performance with async shaders. Afaiu async shaders are just shaders with better scheduling through the ace unit.

As for differences between the 2 brands...if you are in for some reading check out the threads over at beyond3d, I wouldn't know where to start though.

 

[ShintaiDK]

Allegedly the ace units are practically dead silicon running dx11 hardware. That's why Dan Baker of oxide fame has been tweeting about "free" performance with async shaders. Afaiu async shaders are just shaders with better scheduling through the ace unit.

As for differences between the 2 brands...if you are in for some reading check out the threads over at beyond3d, I wouldn't know where to start though.

If we imagine that is the case. Then power consumption will also rise significantly with performance.

Nothing points to any magic waiting for GCN that other uarchs wont have. If anything, it rather points to Maxwellv2 will do even better.

 

[Silverforce11]

If we imagine that is the case. Then power consumption will also rise significantly with performance.

Nothing points to any magic waiting for GCN that other uarchs wont have. If anything, it rather points to Maxwellv2 will do even better.

Why, cos Ryan Smith thinks Maxwell 2 has 32 queue engines?

It's the same as Kepler, 1 queue engine that can do 32 compute or 1 rendering + 1 compute (that Kepler can't do) due to its improved DMA engine.

Different animal to GCN with 1 CP for rendering and 8 ACE queue engines (each with 8 queues).

Then there's the question of 1 crippled DMA engine on consumer SKU with the 2 functional on Teslas. NV has not been clear at all on its uarch.

If Ryan Smith wants to write an indepth on the uarch, he should get his basics facts straight.

http://www.anandtech.com/show/9124/amd-dives-deep-on-asynchronous-shading

Compute Queues should not be confused with Engines. If the table listed represents queues, then GCN 1.1/1.2 onwards should be 1 + 64.

 

[Azix]

If we imagine that is the case. Then power consumption will also rise significantly with performance.

Nothing points to any magic waiting for GCN that other uarchs wont have. If anything, it rather points to Maxwellv2 will do even better.

The silicon isn't turned off I would think. It would not point to maxwell 2 doing better if AMDs chips use more power because they have performance on the table due to silicon using power but not doing useful work. So I don't expect GCN will use more power, but I am not sure.

But that kind of speculation aside, we know Anything below maxwell 2 is lacking against GCN and we know maxwell 2 itself is not quite as parallel as GCN. Surely nothing like the ACEs since maxwell 2 has just one engine.

 

[bystander36]

From reading discussion by gamedevs on this matter, GCN is running crippled on DX11 due to its serial nature, and it needs a very parallel API (Mantle, DX12, Vulkan etc) for it so shine.

We shall see soon enough with the wave of DX12 games incoming.

If this is the case, we should see some big gains on some of the Mantle games out there. Has anyone done some comparisons?

 

[NTMBK]

If this is the case, we should see some big gains on some of the Mantle games out there. Has anyone done some comparisons?

Not all Mantle games use async shaders- as far as I know only Thief does.

 

[Azix]

The best information on async shader performance improvements come from developer interviews and console cases like infamous second son on ps4.

The tomorrow children dev claims 6-10ms saved on a 33ms scene or thereabouts. 18-30% gains. The fable devs claim free graphics and their game looks pre-rendered somewhat.

The two best confirmed benefits of dx12 are lower overhead giving performance boost and asynchronous shaders seen in demos and released games giving very obvious performance boosts. enough for a 290x to approach a 980ti if maxwell 2 does not do well enough there - hilarious if this happens. Fury X wouldnt even be in the same class as a 980ti at that point. Fingers crossed on both sides

Oh, fable legends also answers the question some have asked about when we see dx12 games. Thats a gx12 game to be in beta this summer.

https://youtu.be/7MEgJLvoP2U?t=20m16s

http://gearnuke.com/tomorrow-childr...fits-multiplatform-development-across-ps4xbo/

On a 33.333 ms frame, we saved roughly 6ms by using the async compute pipes. Thats where the 18% comes from. Those savings went up to 10ms on our stress test level. Really if you’re working on a AMD GCN GPU, I can’t recommend looking into async compute enough!

 

[ShintaiDK]

Not all Mantle games use async shaders- as far as I know only Thief does.

If thats the case the benefit is a couple of % at best and pretty much nonexistant.

 

[Azix]

If thats the case the benefit is a couple of % at best and pretty much nonexistant.

Nope. Its a GPU generation's worth of improvement.

I wouldn't base an idea of what performance gains are to be had on thief. Wasn't exactly doing anything fancy.

Depending on what kind of work you are doing and what dependencies it has, its possible to hack it in later in the dev cycle and still get very good results. But like most things in life, you can often get far better results if you plan for it upfront.

You’d be insane not to use it. Two out of the three major gaming platforms use AMD GCN now, so if you’re targeting Xbox and PS4 (sorry WiiU devs!) you can do similar implementations. Supposedly the Xbox One has 4x fewer compute pipes* than PS4, but I would expect that they are totally exposed** to developers and usable.

To all the people who say nvidia is the major player on PC so they will be catered to, don't forget we are still getting console leftovers. We eat what the "peasants" eat. The differences in architecture then start to favor AMD especially when the benefits to be had are that significant.

 

[Azix]

As far as information on the architecture, sites like anandtech do talk about it in detail in some reviews and articles. problem is figuring out what the technical talk leads to in real world performance.

 

[stahlhart]

Stop the arguing and stay on topic.
-- stahlhart

 

[Stuka87]

The best info on this would most likely come from a dev that is not sponsored by either company, but develops highly optimized games for both. Just a matter of some site tracking down said person/team and getting an interview.

 

[Silverforce11]

The best info on this would most likely come from a dev that is not sponsored by either company, but develops highly optimized games for both. Just a matter of some site tracking down said person/team and getting an interview.

Most of the gamedev chatter goes on at b3d & reddit r/gamedev. The is almost zero talk from NV about how good their uarch is for DX12 or more indepth details in that regards. It's really been AMD** that's hyping it up (async compute & VR), but interestingly, also devs have been hyping up GCN for it.

The only reason I'm buying into their hype is because of evidence suggesting the close relationship of Mantle, Vulkan/DX12 & Xbone/PS4 APIs. When devs who work on those engines say lightning, physics, effects as "free performance" or scenes improve in perf by 18-30%, one would have to agree it's got potential.

** One thing I've learnt, AMD hype is unreliable ie. "Overclocker's dream"..