Vega/Navi Rumors (Updated)

[mohit9206]

Pascal is basically a Kepler with new features.

That IS what they're doing. A different name doesn't change the fact that it's iterative.

There hasn't been a clean sheet design from NVIDIA since Tesla. Same for AMD and GCN.

Ok i didn't know that.

 

[zlatan]

A redesign is indicated only when the underlying architecture won't scale anymore. The memory model will be a limiting factor, and to achive good performance, the engineers need to pack a lot of registers and caches to the hardware. A typical example is Intel Gen 9. Which is great if you just see the product, but bad if you compare the size and the performance of the GPU with a similar one from AMD/NV. This is typically an engineering problem. As a customer you don't care about how hard and expensive for Intel to build that hardware, as long as the perf/price ratio is good.

For AMD and NV, a big redisign won't be necessary for the next two rounds, because the main architectures are still scalable enough, and this is also extendable. They need to react some lot bigger problems, like the primitive culling inefficiency, the VRAM management inefficiency, the resource allocation issues (which leads to bad occupancy), the quad efficiency problem, the overdraw issue and the depth information unpredictability. These are the actual problems that will limiting the performance in the future. Vega partially or fully solved a lot of them. And Volta also get some advancements.

 

[CatMerc]

You also tend to do a clean sheet design when you intend to tackle a market you previously didn't. Tesla was NVIDIA's introduction of CUDA, and GCN was introduced when AMD decided they wanted to play in the compute market.

 

[Ajay]

Performance of Vega in current state has little to do with drivers.

Drivers only expose what is available on hardware to the application, if it is designed to use those features. One of reasons for example why Titan Xp got big boost in SpecPerf application with latest driver update was because Nvidia stopper artificially gimping its performance by disabling Antialiased Lines feature in non Quadro hardware. The applications are designed to use this feature, but its drivers job to show it to the application.

Current software is not designed with Programmable Geometry Pipeline and Primitive Shaders in mind.

Classic problem with AMD GFX AIBs as summed up by John Carmack of ID fame: https://twitter.com/id_aa_carmack/status/708681449671495680?lang=en

 

[Stuka87]

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

Just because nVidia comes out with new names does not mean they are new architectures. Because they arent. Everything has been iterative. GCN has also been iterative, as GCN4 varies quite a bit from GCN 1.

 

[Zstream]

Classic problem with AMD GFX AIBs as summed up by John Carmack of ID fame: https://twitter.com/id_aa_carmack/status/708681449671495680?lang=en

Which is mind blowing to me. They have the damn consoles wrapped up and still can't figure it out? I mean come on!

 

[Yakk]

Just because nVidia comes out with new names does not mean they are new architectures. Because they arent. Everything has been iterative. GCN has also been iterative, as GCN4 varies quite a bit from GCN 1.

People buy marketing, new names work better for marketing.

To get onboard with this AMD's would need to come out with new names and just say it's backwards compatible.

 

[Dribble]

Which is mind blowing to me. They have the damn consoles wrapped up and still can't figure it out? I mean come on!

Nvidia has a huge software team, AMD doesn't. It was a choice Nvidia made many years ago - not to just produce hardware and drivers and let others do the rest - but to take an active role in making software for their devices. Hence CUDA, etc. Short term expensive, long term it is a big reason why they are where they are today.

AMD (coming from the background of x86 manufacture where MS makes all the software) took the cheap choice - make the basic drivers and leave the rest to MS/Open Source/anyone. Another short sighted decision by AMD corporate that has got them to where they are today.

 

[Glo.]

Which is mind blowing to me. They have the damn consoles wrapped up and still can't figure it out? I mean come on!

Because Nvidia has very little documentation of their hardware, and does not allow developers do the work, and lets every performance to come from drivers.

AMD designed GCN to give developers control over its performance and has very well documented hardware, but their drivers are not doing a lot of work with the hardware. That is why Developers very often criticize this approach.

 

[Glo.]

Nvidia has a huge software team, AMD doesn't. It was a choice Nvidia made many years ago - not to just produce hardware and drivers and let others do the rest - but to take an active role in making software for their devices. Hence CUDA, etc. Short term expensive, long term it is a big reason why they are where they are today.

AMD (coming from the background of x86 manufacture where MS makes all the software) took the cheap choice - make the basic drivers and leave the rest to MS/Open Source/anyone. Another short sighted decision by AMD corporate that has got them to where they are today.

It isn't short sighted. Drivers alone cannot give you that much, as optimization of application can give you.

Currently most of applications are designed for Nvidia pipelines width(32 KB warp) instead of GCN(64 KB wavefront).

 

[Mopetar]

AMD designed GCN to give developers control over its performance and has very well documented hardware, but their drivers are not doing a lot of work with the hardware. That is why Developers very often criticize this approach.

That's a lot of AMDs problem though. It's hard to get developers on board with spending a lot of time optimizing for a single card that will only be a sliver of the market for several years.

It would be okay if this were going into a console and the game is platform exclusive, but you can't expect developers to spend gobs of time optimizing and tweaking for little payoff from their perspective.

RPM might see more use, but probably only because it's been strongly hinted that NVidia will be implementing it as well. At best Vega makes a good showroom product for AMD to demonstrate where their tech is headed and it may well have legs for the future as those technologies are utilized, but I don't see many existing titles or those being currently developed doing much to take advantage of Vega's theoretical performance.

 

[Glo.]

That's a lot of AMDs problem though. It's hard to get developers on board with spending a lot of time optimizing for a single card that will only be a sliver of the market for several years.

It would be okay if this were going into a console and the game is platform exclusive, but you can't expect developers to spend gobs of time optimizing and tweaking for little payoff from their perspective.

RPM might see more use, but probably only because it's been strongly hinted that NVidia will be implementing it as well. At best Vega makes a good showroom product for AMD to demonstrate where their tech is headed and it may well have legs for the future as those technologies are utilized, but I don't see many existing titles or those being currently developed doing much to take advantage of Vega's theoretical performance.

You do not have to optimize for single card. AMD invested in hardware scheduling to alleviate the need for this. You have to optimize for specific architecture, rather than specific hardware. Thats all.

From developer perspective all you have to program in your application is Geometry Pipeline for Vega, Shader Intrinsics and Primitive Shaders. Rest should be handled by the hardware, if devs will allow it.

 

[Ajay]

It isn't short sighted. Drivers alone cannot give you that much, as optimization of application can give you.
Currently most of applications are designed for Nvidia pipelines width(32 KB warp) instead of GCN(64 KB wavefront).

Well, it turned out to be short sight in so far as Nvidia is winning the graphics sales war right now because of a superior strategy, not necessarily superior hardware. Best intentions don't matter - sales numbers do. So does timing - if Vega RX was out this summer as was expected, I would have considered it along side the GTX 1070, but it still hasn't come out

 

[Phynaz]

From developer perspective all you have to program in your application is Geometry Pipeline for Vega, Shader Intrinsics and Primitive Shaders. Rest should be handled by the hardware, if devs will allow it.

"That's all"?

How many man months do you think a game developer is going to spend on architecture specific enhancements for a tiny fraction of the market? Hint: There's a reason why companies use Gameworks instead of developing their own graphics functions - $$$.

 

[tamz_msc]

The problem with AMD is that it does not command the market share to dictate developers to make optimizations around the features of it's GPU architecture. It's much easier to say that things like programmable geometry pipeline and primitive shaders would make a difference when you're outselling your competitor by 3-to-1 than when you're not.

 

[Magic Hate Ball]

"That's all"?

How many man months do you think a game developer is going to spend on architecture specific enhancements for a tiny fraction of the market? Hint: There's a reason why companies use Gameworks instead of developing their own graphics functions - $$$.

If these functions are available in the Zen-based APUs (as they should if they're Vega based) then it would be a great move for the lower end segment (biggest part of the market) to help your game perform on mass-market systems like iGPU desktops and most laptops?

 

[Glo.]

"That's all"?

How many man months do you think a game developer is going to spend on architecture specific enhancements for a tiny fraction of the market? Hint: There's a reason why companies use Gameworks instead of developing their own graphics functions - $$$.

I am not willing to go into crystal ball and predict future. I am talking about what is required for Vega to show its real potential. You have straight up went into crystal ball.

Secondly. ALL next generation AMD GPUs are based on Vega features. If you believe this is one off architecture, you are more short sighted than you believe. Developers WILL HAVE to optimize for Vega one way or another. Because Primitive Shaders, Programmable Geometry Pipeline, HBCC, are inherent parts not only of GCN5, but all next generation AMD GPUs, for foreseeable future.

If these functions are available in the Zen-based APUs (as they should if they're Vega based) then it would be a great move for the lower end segment (biggest part of the market) to help your game perform on mass-market systems like iGPU desktops and most laptops?

All Vega based GPUs have the same Instruction set, and features.

 

[Phynaz]

Developers WILL HAVE to optimize for Vega one way or another

Why will they have to optimize for Vega when they don't optimize for AMD's current GPU architecture? Again - Gameworks.

 

[tamz_msc]

Developers WILL HAVE to optimize for Vega one way or another

Why will they have to optimize for Vega when they don't optimize for AMD's current GPU architecture? Again - Gameworks.

You make it sound like using Gameworks by itself makes a game well-optimized on NVIDIA hardware, but that's simply not true. Developers optimize for NVIDIA because they have the majority market share by a significant margin.

 

[Glo.]

Why will they have to optimize for Vega when they don't optimize for AMD's current GPU architecture? Again - Gameworks.

GameWorks is forced when Nvidia is paying the implementation development of it. And it hurts both vendors: AMD and Nvidia, because as a result you get worse performance, than you could get without it.

Funniest part. Primitive Shaders, and Programmable Geometry Pipeline are designed to get rid of gameWorks performance problems on AMD hardware.

You make it sound like using Gameworks by itself makes a game well-optimized on NVIDIA hardware, but that's simply not true. Developers optimize for NVIDIA because they have the majority market share by a significant margin.

Developers optimize for GameWorks, because they are payed for this, by Nvidia.

Programmable Geometry Pipeline and Primitive Shaders simplifies the development process, because you bypass few key stages with them, so it is essentially saving time and effort, and in the end - saves money for the developers. I have a hard time, why you guys see problem in this feature, where there is none.

 

[tamz_msc]

Developers optimize for GameWorks, because they are payed for this, by Nvidia.

I find the reasons much less sinister - it's simply due to market share. Out of a dozen or so AAA titles that are released every year, hardly two or three use Gameworks. This talk about Gameworks is a huge digression. Let's keep the discussion on topic.

I'm still maintaining that AMD doesn't have the traction to dictate development for it's technologies, which is why they need to deliver on things like improved geometry performance, with minimal effort from developers. Somebody has to do non-trivial programming to make the "programmable" part of the programmable geometry pipeline work. That's already asking a lot, given the state of development on low-level APIs.

 

[Elixer]

Navi will be just slightly updated GCN5 architecture. Vega is foundation for all next generation GCN architectures.

We don't know that for sure.

Leapfrogging design teams suggests they are working on multiple products at once, and Navi could be a new design, much like the way Ryzen was a new design from bulldozer.

Heck, they need at least 2x, if not 3x(4x?) the performance of Vega to compete at the very high end (Volta), and you don't do that by only doing incremental updates.

 

[Ajay]

Again, in the side above, the elusive 14nm+ shows up, sure would be nice to get something official about what this node is.

 

[Phynaz]

You make it sound like using Gameworks by itself makes a game well-optimized on NVIDIA hardware, but that's simply not true. Developers optimize for NVIDIA because they have the majority market share by a significant margin.

That's my point. It's not optimized, it's inexpensive to implement. It's a plug and play black box.

 

[Phynaz]

[QUOTE="Glo., post: 39025628, member: 359189"
Programmable Geometry Pipeline and Primitive Shaders simplifies the development process, because you bypass few key stages with them, so it is essentially saving time and effort, and in the end - saves money for the developers.[/QUOTE]

This is ridiculous. It requires developer time to implement architecture specific features.

What are these "stages" that developers "bypass"?