AMD VEGA confirmed for 2017 H1

[D.P.]

Maybe i am crazy but did AMD not co-develop HBM with Hynix? Surely if this were true, they have some kind of priority clauses in place for something they contributed to? And i cannot imagine they will be paying the same price as other purchasers for the product they helped develop surely?

As the product is also being developed in multiple manufacturing facilities from more than 1 manufacturer i do not expect there to be a huge shortage for AMD for HBM2, what is to say they have no contract in place for preferential supply for a product they helped develop and at a cheaper price?

Too many people read the nonsense that is on such sites like WCCFTech without actually stopping for a second, and thinking, and then coming to forums like this and posting said nonsense, without doing even the smallest bit of research to back up the claims.

HBM is JEDEC industry standard so anyone can develop it and AMD has no priority or preferential treatment. They may or may no have some contract with Hynix but Hynix is behind Samsung and quite clearly AMD have no special clause with Samsung as Nvidia are getting the majority of chips. IF AMd arranged a special contract with Hynix it may actaully be backfiring with AMD forced to use Hynix ans its HBM supplier but Hynix is not yet able to deliver.

And just because there is 2 suppliers does not mean there is sufficient supply, quite the opposite since Hynix quite clearly is unable to supply currently. HBM is complex and expensive to produce, the total market is relatively small so production facilities are not set u cope with very high volume production.

 

[TNA]

Too many people read the nonsense that is on such sites like WCCFTech without actually stopping for a second, and thinking, and then coming to forums like this and posting said nonsense, without doing even the smallest bit of research to back up the claims.

LoadsaMoney comes to mind

 

[Rroff]

AMD designed their own Interposer, In partnership with Hynix they also designed the HBM architecture, they own IP in it.

Hence my point that WCCF tech might have just used the price available to the general public rather than the pricing big companies would be able to buy at.

 

[LoadsaMoney]

LoadsaMoney comes to mind

 

[N19h7m4r3]

Vega just needs to beat the GTX 1060, and Loadsa will be happy, and buy 2 for Crossfire.

 

[humbug]

Vega just needs to beat the GTX 1060, and Loadsa will be happy, and buy 2 for Crossfire.

Speaking of Cross Fire, i wonder how AMD's Infinity Fabric will effect that, if at all, on the CPU side it allows for multiple CPU's on a motherboard at 100% scaling, and Multiple CPU's on a single PCB.

Vega GPU's will have Infinity Fabric, there must be a reason for it, maybe as it does with CPU's its a way of having Cross Fire scaling natively, IE two, three or four GPU's be it on the same PCB or separate GPU's would scale without any input from software or games.

Edit, it may also benefit AMD, smaller GPU yield much better than larger ones, so AMD might Make Vega with bunches of smaller 'GPU'lets' instead of one big one, the higher up the GPU the more GPU'lets it has.
Theoretically the Graphics Cards performance is limited only by PCB power consumption, this other than how many GPU'lets you can fit on a PCD.

This is what they are already doing with their 16, 32 and 48 core ZEN CPU's.

Spoiler: 4 CPU's on a PCB

 

[N19h7m4r3]

Speaking of Cross Fire, i wonder how AMD's Infinity Fabric will effect that, if at all, on the CPU side it allows for multiple CPU's on a motherboard at 100% scaling, and Multiple CPU's on a single PCB.

Vega GPU's will have Infinity Fabric, there must be a reason for it, maybe as it does with CPU's its a way of having Cross Fire scaling natively, IE two, three or four GPU's be it on the same PCB or separate GPU's would scale without any input from software or games.

Until Implicit/Explicit Multi-adapter becomes common, along with DX12/Vulkan it'll be a long time before we get similar effects in GPUs.

Raja has stated in 2016 that their end goal is to have Multi GPU be as common as Multicore CPUs, but the APIs need to be properly adopted, and developers need to get on it.

It's simply can't be done well enough under DX11, and needs massive driver support

Also the IF on Vega is for instinct to connect better with Naples as well, and to be the ground work for Navi.

 

[AllBodies]

Edit, it may also benefit AMD, smaller GPU yield much better than larger ones, so AMD might Make Vega with bunches of smaller 'GPU'lets' instead of one big one, the higher up the GPU the more GPU'lets it has.
Theoretically the Graphics Cards performance is limited only by PCB power consumption, this other than how many GPU'lets you can fit on a PCD.

This is what they are already doing with their 16, 32 and 48 core ZEN CPU's.

Spoiler: 4 CPU's on a PCB

There's a lot of speculation this is what Navi will be. But Vega was designed before the tech was finished, so will be traditional.

Combining 4 200mm2 GPUs would be both hilariously powerful and hilariously good for yields. So if it's possible for them to do it, they will try.

 

[Rroff]

Infinity Fabric. etc. will almost certainly be about scaling compute performance with big data sets, etc. as N19h7m4r3 said we need explicit multi-adapter support to mature for the best possible utilisation of multiple GPUs with current GPU architectures.

 

[Gregster]

I'm not in the least bit surprised

You still looking to buy in Q4 yourself? Or are you going to jump sooner?

 

[humbug]

Until Implicit/Explicit Multi-adapter becomes common, along with DX12/Vulkan it'll be a long time before we get similar effects in GPUs.

Raja has stated in 2016 that their end goal is to have Multi GPU be as common as Multicore CPUs, but the APIs need to be properly adopted, and developers need to get on it.

It's simply can't be done well enough under DX11, and needs massive driver support

Also the IF on Vega is for instinct to connect better with Naples as well, and to be the ground work for Navi.

There's a lot of speculation this is what Navi will be. But Vega was designed before the tech was finished, so will be traditional.

Combining 4 200mm2 GPUs would be both hilariously powerful and hilariously good for yields. So if it's possible for them to do it, they will try.

Navi does make more sense, i agree Vega is perhaps too early.

@N19h7m4r3 the whole point of Infinity Fabric is for multiple chips to act as one piece of hardware, the game or API doesn't see 2 or 4 GPU's, it just sees one because through hardware they are connected together and act as one.

Anyone with a Ryzen Processor actually has two CPU's, under the DIE cap there are two 4 core CPU's connected together through the Infinity Fabric, it operates through the system memory.
Its actually an evolution of AMD's Heterogeneous System Architecture (HSA) that connected the CPU PFU with the iGPU's FPU in a similar way, so the CPU's X86 executions were able to use the GPU's much faster FPU, the result was.....

Spoiler: Huge performance boost.

 

[Panos]

Speaking of Cross Fire, i wonder how AMD's Infinity Fabric will effect that, if at all, on the CPU side it allows for multiple CPU's on a motherboard at 100% scaling, and Multiple CPU's on a single PCB.

Vega GPU's will have Infinity Fabric, there must be a reason for it, maybe as it does with CPU's its a way of having Cross Fire scaling natively, IE two, three or four GPU's be it on the same PCB or separate GPU's would scale without any input from software or games.

Edit, it may also benefit AMD, smaller GPU yield much better than larger ones, so AMD might Make Vega with bunches of smaller 'GPU'lets' instead of one big one, the higher up the GPU the more GPU'lets it has.
Theoretically the Graphics Cards performance is limited only by PCB power consumption, this other than how many GPU'lets you can fit on a PCD.

This is what they are already doing with their 16, 32 and 48 core ZEN CPU's.

Spoiler: 4 CPU's on a PCB

Something raised many pages ago, based on what Raja said in the event back in Feb/March

a) The system would handle multiple GPUs as a single entity, and was about the Vega design. Nothing to do with Crossfire.
b) A system with Ryzen CPU and a Vega GPU would form an ecosystem. And is possible because the Ryzen CPUs are SOC and they have direct communication with the graphic cards.

All this tech is not that far away from the current gen of consoles. Yes an Intel CPU would work as "normal" with a Vega GPU. So does a NV GPUs (albeit with crappy drivers that affect the Ryzen system performance) with the Ryzen CPUs. But if you put them all together, we might have something more powerful.

Hell we do not know how Infinity fabric would work with CPU+GPU in this case. Might see the HBM2 as cache as is Raja said back in January CES, and improve the performance of all the components. Not having something in the middle in the form of chipset, AMD could pull off anything.

And lets not forget, Infinity Fabric is the great unknown and a new tech.
On OCNET some who tested, against the general advice, dual sided Samsung-B die, saw at 2666C12 (3600 rated but couldn't OC higher three weeks ago) the same performance in games as single sided Samsung-B at 3200C14 and 15% higher than single sided 2666C12. (which was 8% higher than 2133).

And that 16 core 32 thread monstrosity, having quad channel ram bandwidth available, could be in for surprising us all.

 

[N19h7m4r3]

Navi does make more sense, i agree Vega is perhaps too early.

@N19h7m4r3 the whole point of Infinity Fabric is for multiple chips to act as one piece of hardware, the game or API doesn't see 2 or 4 GPU's, it just sees one because through hardware they are connected together and act as one.

Anyone with a Ryzen Processor actually has two CPU's, under the DIE cap there are two 4 core CPU's connected together through the Infinity Fabric, it operates through the system memory.
Its actually an evolution of AMD's Heterogeneous System Architecture (HSA) that connected the CPU PFU with the iGPU's FPU in a similar way, so the CPU's X86 executions were able to use the GPU's much faster FPU, the result was.....

Even with IF, any game will still see two GPUs, just like they will see 4-8 different CPU cores. The developers still need to put the effort into making it all work.

Infinity Fabric is still a very high speed interconnect that makes communication easier. HBCC + HBM2 seems essential along with I.F to get to the point that AMD wants, but without developer support it won't work.
Just look at the early issues on Ryzen's launch with NUMA issues, and some games not even knowing how to handle the CCXs correctly.

HSA also needed dedicated developer work. One thing that should help is AMD partnering with big publisher/developer houses; and getting it sorted on the enterprise side.

I can see why AMD is purely banking on HBM2 for the future, and call it a cache as well. Mix that with the HBCC, and Infinity Fabric and we might get the future they want.

 

[humbug]

Even with IF, any game will still see two GPUs, just like they will see 4-8 different CPU cores. The developers still need to put the effort into making it all work.

Infinity Fabric is still a very high speed interconnect that makes communication easier. HBCC + HBM2 seems essential along with I.F to get to the point that AMD wants, but without developer support it won't work.
Just look at the early issues on Ryzen's launch with NUMA issues, and some games not even knowing how to handle the CCXs correctly.

HSA also needed dedicated developer work. One thing that should help is AMD partnering with big publisher/developer houses; and getting it sorted on the enterprise side.

I can see why AMD is purely banking on HBM2 for the future, and call it a cache as well. Mix that with the HBCC, and Infinity Fabric and we might get the future they want.

I don't know enough about it to explain it, but what you are citing is basically a Cross Fire bridge, multiple GPU's working on separate code paths.

Thats not what Infinity Fabric is, read above you, Infinity Fabric is Multiple Hardware, be it CPU's or GPU's as one, your 1800X 'if you had one; Windows does not see two 4 core CPU's, it sees one 8 core CPU.

It does not see this as four 8 core CPU's, actually just as an 1800X is two 4 cores that is eight 4 core CPU's.... but it sees it as one 32 core CPU.

Be it Vega or Navi as Raja explained is lots of GPU's acting as one, no Cross Fire, Cross Fire no longer Exists, its defunct, dead. The software, the API just sees one GPU no matter how many there actually are, just as Windows sees one CPU, not two, Four or six or however many more AMD will stuff onto a PCB.

 

[D.P.]

Speaking of Cross Fire, i wonder how AMD's Infinity Fabric will effect that, if at all, on the CPU side it allows for multiple CPU's on a motherboard at 100% scaling, and Multiple CPU's on a single PCB.

Vega GPU's will have Infinity Fabric, there must be a reason for it, maybe as it does with CPU's its a way of having Cross Fire scaling natively, IE two, three or four GPU's be it on the same PCB or separate GPU's would scale without any input from software or games.

Edit, it may also benefit AMD, smaller GPU yield much better than larger ones, so AMD might Make Vega with bunches of smaller 'GPU'lets' instead of one big one, the higher up the GPU the more GPU'lets it has.
Theoretically the Graphics Cards performance is limited only by PCB power consumption, this other than how many GPU'lets you can fit on a PCD.

This is what they are already doing with their 16, 32 and 48 core ZEN CPU's.

There is a big difference between a CPU and a GPU though.


With CPU cores they don't magically wall work together. It takes a lot of work by developers to create appropriately multi-threaded software. All the standard, tools, libraries and language constructs simply didn't even work in a multi-threaded way. Furthermore, many takes on a CPU are data independent so there are no benefits form cache coherence and such like between threads and so there is really no difference if 2 threadare running on 2 cores on the same die, on different dies, on different chips (and in some cases, different systems). Graphics is not like that, e.g. rendering a landscape in a game with have shared textures, shaders, lighting data etc and the cores can make much better use of shared cache.



Infinity Fanric doesn't magically make a GPU appear as a single device, it is just higher bandwidth and lower latency cross-fire. AMD's equivalent to NVlink.


While I agree that utilizing multiple smaller GPU dies on some kind of interposer with shared memory has many benefits, not even Navi will be bringing that to the table in a seamless fashion for developers. To get all GPU cores working efficiently will require complex load balancing and the only way to truly get optimal performance would be explicit coding from the developer. There is a reason both AMD and Nvidia dropped split frame rendering for Alternate frame rendering and those issues apply just as equally with Infinity fabric.

 

[humbug]

There is a big difference between a CPU and a GPU though.


With CPU cores they don't magically wall work together. It takes a lot of work by developers to create appropriately multi-threaded software. All the standard, tools, libraries and language constructs simply didn't even work in a multi-threaded way. Furthermore, many takes on a CPU are data independent so there are no benefits form cache coherence and such like between threads and so there is really no difference if 2 threadare running on 2 cores on the same die, on different dies, on different chips (and in some cases, different systems). Graphics is not like that, e.g. rendering a landscape in a game with have shared textures, shaders, lighting data etc and the cores can make much better use of shared cache.



Infinity Fanric doesn't magically make a GPU appear as a single device, it is just higher bandwidth and lower latency cross-fire. AMD's equivalent to NVlink.


While I agree that utilizing multiple smaller GPU dies on some kind of interposer with shared memory has many benefits, not even Navi will be bringing that to the table in a seamless fashion for developers. To get all GPU cores working efficiently will require complex load balancing and the only way to truly get optimal performance would be explicit coding from the developer. There is a reason both AMD and Nvidia dropped split frame rendering for Alternate frame rendering and those issues apply just as equally with Infinity fabric.

No one said it was easy. but besides that, Raja, who lets face it as AMD chief architect knows more about Infinity Fabric than you ever will... it seems disagrees with you, tell him he doesn't understand his own invention, its nothing to do with me.

 

[LoadsaMoney]

Vega just needs to beat the GTX 1060, and Loadsa will be happy, and buy 2 for Crossfire.

2x would only beat it in the odd game that xfire worked in without issue

 

[D.P.]

I don't know enough about it to explain it, but what you are citing is basically a Cross Fire bridge, multiple GPU's working on separate code paths.

Thats not what Infinity Fabric is, read above you, Infinity Fabric is Multiple Hardware, be it CPU's or GPU's as one, your 1800X 'if you had one; Windows does not see two 4 core CPU's, it sees one 8 core CPU.

It does not see this as four 8 core CPU's, actually just as an 1800X is two 4 cores that is eight 4 core CPU's.... but it sees it as one 32 core CPU.

Be it Vega or Navi as Raja explained is lots of GPU's acting as one, no Cross Fire, Cross Fire no longer Exists, its defunct, dead. The software, the API just sees one GPU no matter how many there actually are, just as Windows sees one CPU, not two, Four or six or however many more AMD will stuff onto a PCB.

You really don't seem to understand this.


Yous aid it your self, with infinti Fabric the OS sees an 8 core CPU. Now if Infinity Fabric was really doing something magically then the OS would just see 1 core, and all kidn fo complex magic would happen that would automatically split a task between 8 different cores giving an 8x speed up. In relaity that doesn't exist in the slightest. You give a task to be computed and it will get done by 1 single core. If you want to use more than 2 then you will have to code for that explicitly and decide what work each core is going to do.


With a GPU it is tasked with rendering the entire scene in 1 frame. It already has thousands of cores, and will be distributing the work between the. This is where some of the inefficiency in recent AMD GPUs have come from because it is not an easy task to keep all those cores balanced and busy. Now f you have 2 or more GPUs 1 of those, or some additional controller like the CPU has to decide what work each GPU is going to do. the naive solution is just to split the frame in 2, butt the problem is the work loads wont be balanced and you will never get decent scaling. Thus it will need explicit coding form the software to decide how best to distribute the workload.

 

[D.P.]

No one said it was easy. but besides that, Raja, who lets face it as AMD chief architect knows more about Infinity Fabric than you ever will... it seems disagrees with you, tell him he doesn't understand his own invention, its nothing to do with me.

You just don't even understand what Raja is saying, and your displayed knowledge of CPU mutli-threading would indicate you have a lot of learning to do before you start trying to educate others.

 

[humbug]

Oh stop with the "i'm smarter than you" ^^^^^ you're not, you use that card with everyone who disagrees with you, its getting old and tiresome.
There is nothing to be misunderstood in "multiple GPU's acting as one" get real. no matter how complex you try and make it to then claim intellectual arguments won by trying to bamboozle people; what he said was utterly simple.

You talk nothing but gibberish and claim its because you're smarter than everyone, when in actual fact the subject and what was said is very simple.

And yes i know enough about it to know gibberish when i see it.

You really don't seem to understand this.


Yous aid it your self, with infinti Fabric the OS sees an 8 core CPU. Now if Infinity Fabric was really doing something magically then the OS would just see 1 core, and all kidn fo complex magic would happen that would automatically split a task between 8 different cores giving an 8x speed up. In relaity that doesn't exist in the slightest. You give a task to be computed and it will get done by 1 single core. If you want to use more than 2 then you will have to code for that explicitly and decide what work each core is going to do.


With a GPU it is tasked with rendering the entire scene in 1 frame. It already has thousands of cores, and will be distributing the work between the. This is where some of the inefficiency in recent AMD GPUs have come from because it is not an easy task to keep all those cores balanced and busy. Now f you have 2 or more GPUs 1 of those, or some additional controller like the CPU has to decide what work each GPU is going to do. the naive solution is just to split the frame in 2, butt the problem is the work loads wont be balanced and you will never get decent scaling. Thus it will need explicit coding form the software to decide how best to distribute the workload.

Same thing, so again...

No one said it was easy. but besides that, Raja, who lets face it as AMD chief architect knows more about Infinity Fabric than you ever will... it seems disagrees with you, tell him he doesn't understand his own invention, its nothing to do with me.