[WCCF] AMD Radeon R9 390X Pictured

[DrMrLordX]

That is the functional base of Mantle, that has been used by developers of Vulkan and DX12, and Metal. It also allowes to use system memory for GPU purpose as well.

You should know know what that means . RAM for big Data, HBM for data that requires gigantic bandwith and low latency. Also helpful for systems with unified memory, eg: consoles, CPUs with integrated Graphics, smartphones...

Sorry for off-top. Back to topic, please .

In all fairness to you, that is somewhat topical. It could explain why AMD is confident that 4Gb of HBM VRAM is enough for the 390X if that is, in fact, the maximum amount available at launch.

That image. hrnnnnnnggggggg!!

On the card size/looks issue, I remember discussing this with the GF when I said I thought it was ugly. I used the metaphor of muscle cars to sports cars. I prefer the look of classic muscle cars over some of the newer stream line sports cars. It's just aesthetics and ever one has a preference.

Chicks dig the long card, m i rite?

I could see major engines devoting the time.
You don't have to remake the multi-gpu wheel for each game.

Exactly. DX12 and Vulkan will offer low-level features that big dev shops will bake into their engines. Small shops will be less-inclined to do their own engine in-house when they can use more-sophisticated tools available on license from the big boys. Also, with all the APUs out there from Intel AND AMD, vendor-agnostic multigpu will be just what the doctor ordered, at least for stuff like Kaveri, Carrizo, Broadwell, Skylake, etc.

 

[Tuna-Fish]

I could see major engines devoting the time.
You don't have to remake the multi-gpu wheel for each game.

You have to statically partition your rendering workload. This is is something you do need to do for every game.

The only reasonable target for multi-gpu in the future is having a gpu per eye on VR. I absolutely don't think that most DX12/Vulkan games will be able to make good use of multi-gpu in ways other than afr.

 

[Glo.]

There is a solution, to this Tuna-Fish, if we stay off topic, and discuss technicalities.

I have heard that somebody in industry thinked: is there a way to "bind" two seperate GPUs to be seen by the system/application as one big unit. It would be extremely useful for heterogeneous computing with asynchronous capabilities. In which AMD excells. And guess why people thinked that way? Because for that would be needed an API of some sort...

Also for situations where you have AMD APU with AMD dGPU...

At least for sure, that was discussed inside AMD some time ago.

 

[el etro]

Charlie at semiaccurate suggests AMD could use HBM on the cards not to make the cards faster, but instead to make them more power efficient.

I bet he knows something.

And would be cool to have a HBM GCN card late this year...

 

[Kenmitch]

There is a solution, to this Tuna-Fish, if we stay off topic, and discuss technicalities.

I have heard that somebody in industry thinked: is there a way to "bind" two seperate GPUs to be seen by the system/application as one big unit. It would be extremely useful for heterogeneous computing with asynchronous capabilities. In which AMD excells. And guess why people thinked that way? Because for that would be needed an API of some sort...

Also for situations where you have AMD APU with AMD dGPU...

At least for sure, that was discussed inside AMD some time ago.

Might work? Would be interesting.

Probably next update of GW would include some silly code to limit shaders.

 

[Serandur]

Charlie at semiaccurate suggests AMD could use HBM on the cards not to make the cards faster, but instead to make them more power efficient.

I bet he knows something.

And would be cool to have a HBM GCN card late this year...

That's been fairly evident. One of HBM's majorly advertised features is a reduction in memory power consumption (low clock speeds on the memory being more than made up for by the additional memory channels) by 40-50% and coupled with the lack of a process node shrink and the lack of a new GCN 2.0 architecture, it's made a lot of sense to look at gen 1 HBM as AMD's solution to releasing a 28nm 390X that doesn't put out ridiculous amounts of heat yet can still vastly outperform the 290X with the side benefits of reducing PCB/GPU complexity and providing additional bandwidth.

 

[3DVagabond]

There is a solution, to this Tuna-Fish, if we stay off topic, and discuss technicalities.

I have heard that somebody in industry thinked: is there a way to "bind" two seperate GPUs to be seen by the system/application as one big unit. It would be extremely useful for heterogeneous computing with asynchronous capabilities. In which AMD excells. And guess why people thinked that way? Because for that would be needed an API of some sort...

Also for situations where you have AMD APU with AMD dGPU...

At least for sure, that was discussed inside AMD some time ago.

FWIU, and I'm no expert, the issue is getting the timing right with all of those threads. ~8000 threads, for example, with 2x Fiji, working on current leaked specs.

 

[dacostafilipe]

FWIU, and I'm no expert, the issue is getting the timing right with all of those threads. ~8000 threads, for example, with 2x Fiji, working on current leaked specs.

And those caches ... Won't somebody please think of the caches!

:whiste:

 

[DarkKnightDude]

Couldn't they release a 4GB first then release an 8GB version later? Didn't 8GB 290Xs not show till recently?

 

[ShintaiDK]

Couldn't they release a 4GB first then release an 8GB version later? Didn't 8GB 290Xs not show till recently?

They could if Hynix makes a bigger HBM1 module later. But that may not happen if they put all focus on HBM2.

 

[3DVagabond]

They could if Hynix makes a bigger HBM1 module later. But that may not happen if they put all focus on HBM2.

They don't need a bigger HBM1 module.

AMD manages to integrate 8GB of HBM memory at its peak thanks to a series of improvements to the interposer TSV developed with SK Hynix: the technology in question is called "Dual Link Interposer" (from 4 to 8-Hi-Hi). With a Dual Link Interposing design, SK Hynix will be able to stack 4x (Dual 1GB HBM modules) via an Interposer (2.5D stacking).This will allow AMD to take advantage of a larger amount of memory without having to wait for the arrival of the HBM 2nd generation.

2x 4-HI HBM1 (which should technically be called 8-Hi-Hi according to nomenclature rules) features a 1024-bit interface, two prefetch operations per IO (dual command) and can push 128GB per second per pin. The tRC is 48nm, tCCD is 2ns (1tCK), and VDD is 1.2V.

 

[ShintaiDK]

They don't need a bigger HBM1 module.

Feel free to show me where on Hynix site you find this. because when I search there is nothing on Hynix site refering to "Dual Link Interposer".

Until then it just smells like another forum creation without hold in reality.

 

[Tuna-Fish]

There is a solution, to this Tuna-Fish, if we stay off topic, and discuss technicalities.

I have heard that somebody in industry thinked: is there a way to "bind" two seperate GPUs to be seen by the system/application as one big unit. It would be extremely useful for heterogeneous computing with asynchronous capabilities. In which AMD excells. And guess why people thinked that way? Because for that would be needed an API of some sort...

Making two gpus look like one in software is exactly what the older APIs did. It also makes the kind of low-level optimization like seen in DX12/Mantle/Vulkan mostly impossible. The whole point of these new APIs is to move from an abstract model where programmers utilize a black-box pipeline that can do anything behind the covers towards a model where the programmer can see everything that happens.

The new APIs do have a way to use AFR like the older ones, but with all the same drawbacks. Beyond that, they allow partitioning your graphics load across your gpus any way you like, which would allow efficient use of differently powered gpus, or even gpus from two different manufacturers, but the programmers have to actually do that, and they have to plan in advance for every different configuration they want to support.

 

[dacostafilipe]

Feel free to show me where on Hynix site you find this. because when I search there is nothing on Hynix site refering to "Dual Link Interposer".

Until then it just smells like another forum creation without hold in reality.

Feel free to show us where on Hynix site you find the max 4GB addressable space limit ...

 

[ShintaiDK]

Feel free to show us where on Hynix site you find the max 4GB addressable space limit ...

You mean in relation to shipping HBM1 densities.

Easy:
https://www.skhynix.com/products/support/databook.jsp

But it shouldnt be a surprise:

 

[raghu78]

You mean in relation to shipping HBM1 densities.

Easy:
https://www.skhynix.com/products/support/databook.jsp

But it shouldnt be a surprise:

1 GB using a 4 Hi HBM stack has been known for a long time. What you and many others like you have failed to prove is there can only be 4 such 4 Hi HBM stacks on an interposer beside a GPU. :whiste:

 

[dacostafilipe]

1 GB using a 4 Hi HBM stack has been known for a long time. What you and many others like you have failed to prove is there can only be 4 such HBM stacks on an interposer beside a GPU. :whiste:

This.

 

[exar333]

This.

Generally leaks and proof are to support products that DO exist, or maybe have been cancelled. What information or proof, besides dreams and hopes, do you have that a 8GB card exists?

I HOPE a 8GB card is released, but assuming it will be because no one can disprove it, is silly.

Using your train of thought, a 12/16GB card must be released too, since no one has disproved 3 or 4 stacks as well? Complete silliness.

 

[Kenmitch]

Generally leaks and proof are to support products that DO exist, or maybe have been cancelled. What information or proof, besides dreams and hopes, do you have that a 8GB card exists?

I HOPE a 8GB card is released, but assuming it will be because no one can disprove it, is silly.

Using your train of thought, a 12/16GB card must be released too, since no one has disproved 3 or 4 stacks as well? Complete silliness.

Nor do we even know how HBM performs. 4GB might outperform 8GB's of GDDR5 with enough GPU grunt to back it up. Could be fast and efficient enough to not really matter.

Taking the same GPU for example which is faster? The one with 1GB's of GDDR5 or the one with 2GB's of GDDR3? Past history shows when switching to new memory architecture the amount of memory isn't the deciding factor as far as performance goes.

8GB's does sound better on paper and for marketing reasons.

 

[ShintaiDK]

Nor do we even know how HBM performs. 4GB might outperform 8GB's of GDDR5 with enough GPU grunt to back it up. Could be fast and efficient enough to not really matter.

Taking the same GPU for example which is faster? The one with 1GB's of GDDR5 or the one with 2GB's of GDDR3? Past history shows when switching to new memory architecture the amount of memory isn't the deciding factor as far as performance goes.

8GB's does sound better on paper and for marketing reasons.

The problem is that if you need 1½GB. Then the 1GB card in your case will crawl.

 

[maddie]

Generally leaks and proof are to support products that DO exist, or maybe have been cancelled. What information or proof, besides dreams and hopes, do you have that a 8GB card exists?

I HOPE a 8GB card is released, but assuming it will be because no one can disprove it, is silly.

Using your train of thought, a 12/16GB card must be released too, since no one has disproved 3 or 4 stacks as well? Complete silliness.

Really, I thought that a 4GB card would have 4 stacks?

 

[dacostafilipe]

Generally leaks and proof are to support products that DO exist, or maybe have been cancelled. What information or proof, besides dreams and hopes, do you have that a 8GB card exists?

I HOPE a 8GB card is released, but assuming it will be because no one can disprove it, is silly.

Using your train of thought, a 12/16GB card must be released too, since no one has disproved 3 or 4 stacks as well? Complete silliness.

It's not about "knowing" if there's going to be a 8GB,12GB,etc 390X card. It's about "knowing" if it's possible or not.

If there was some confirmation that HBM1 would have a max 4GB addressable space, we could rule out the 8GB. But until now, there's no such think and that's why I see no reason NOT to expect a 8GB card.

 

[LTC8K6]

This 4gb/8gb argument will be permanently ended soon.

Let's just calm down and wait.

 

[railven]

This 4gb/8gb argument will be permanently ended soon.

Let's just calm down and wait.

But then what's the point of a forum?

 

[boozzer]

This 4gb/8gb argument will be permanently ended soon.

Let's just calm down and wait.

hehe, the mud slinging targeted at amd is in full swing. so don't be surprise if there are more of it towards launch, not less.