AMD 14nm FinFET Polaris GPU Size Leaks Out – 232mm² Large Die

[Citadelen]

30 minutes ago, Mihle Gaming said:

is it that hard to just take their top end card where they do care about power consumption and put in more transistors to get it to perform better? (without a complete re design)

Efficiency is the main aim for Polaris, Pascal will probably be faster but will be hotter and consume more energy.

[Sakkura]

On 16/2/2016 at 4:59 PM, DocSwag said:

In reality though they called the 7970 gcn 1.0, 290x gcn 1.1, and fury x 1.2.

It makes sense to call it 4th gen because it is the fourth iteration of GCN.

Looking at the above it would make sense to be 1.3, but this is the biggest change they have done to their GPUs since GCN itself so it makes sense to be 2.0.

They call it 4th generation GCN, not GCN 4.0. It still isn't certain exactly what it will be called though, so you could be right or I could. Only time will tell

No they didn't. The press came up with those names, not AMD. But "4th generation GCN" is an official name. Don't need time to tell. 

[patrickjp93]

3 hours ago, Mihle Gaming said:

is it that hard to just take their top end card where they do care about power consumption and put in more transistors to get it to perform better? (without a complete re design)

Architectures do come with their own built-in scaling limits for adding resources. Fiji was one such example where 128 ROPs and TMUs were needed, but GCN had a hard limit. Yes. If you think it can be done, go get an electrical engineering degree and turn the industry on its ear.

[Prysin]

1 hour ago, patrickjp93 said:

Architectures do come with their own built-in scaling limits for adding resources. Fiji was one such example where 128 ROPs and TMUs were needed, but GCN had a hard limit. Yes. If you think it can be done, go get an electrical engineering degree and turn the industry on its ear.

well, in Fijis defence, it was probably built for TSCM 20nm.... but that blew up so they had to cut it back in some regards whilst trying to retain performance.

I still think they got too many shaders compared to ROP and TMUs, but i got no idea whether they COULD have added more of those without making it a worse catastrophy then it ended up being

[Mihle]

27 minutes ago, patrickjp93 said:

Architectures do come with their own built-in scaling limits for adding resources. Fiji was one such example where 128 ROPs and TMUs were needed, but GCN had a hard limit. Yes. If you think it can be done, go get an electrical engineering degree and turn the industry on its ear.

I dont know how stuff like that works, thats why I am asking, I am curious.

[patrickjp93]

1 hour ago, Mihle Gaming said:

I dont know how stuff like that works, thats why I am asking, I am curious.

Long story short, it is difficult, or we'd stay with the same architecture with just mild tweaks each year.

[Mihle]

22 minutes ago, patrickjp93 said:

Long story short, it is difficult, or we'd stay with the same architecture with just mild tweaks each year.

ok

[Jahramika]

On 2/14/2016 at 8:38 AM, stconquest said:

Is Polaris introducing CUDA support on AMD?

Yes

[Jahramika]

On 2/14/2016 at 1:55 PM, -BirdiE- said:

Implying 100% of a card's size is related to the manufacturing size...

 

That's why graphics cards have been getting way smaller over the years.... OH wait..

No that's because of Amd's Hbm memory design.

[Jahramika]

21 hours ago, patrickjp93 said:

Your wants are low down on the list compared to the needs of data centers. Better get used to that idea.

They can release cards to address data centers,  base users and fan boys alike

[Notional]

31 minutes ago, Jahramika said:

Yes

What are you talking about? There is less than 0% chance that Polaris will have CUDA support.

[Big-G]

10 hours ago, Citadelen said:

...Pascal will probably be faster but will be hotter and consume more energy.

Conjecture, any backing to that?

[-BirdiE-]

34 minutes ago, Jahramika said:

No that's because of Amd's Hbm memory design.

Not even sure what you're getting at man.

 

I was saying that, with Polaris, just because the manufacturing process is half the size does not mean you will get a GPU that's equally as powerful as the R9 Nano at half the width. Do you disagree with this?

[Trixanity]

10 minutes ago, Notional said:

What are you talking about? There is less than 0% chance that Polaris will have CUDA support.

I think he's referencing the AMD initiative to convert CUDA code to openCL. So it's kinda a yes and no to that. AMD cards will never support CUDA, however if CUDA applications can easily be ported to openCL, they'll have a reasonably similar outcome as if they were natively supporting CUDA (I guess).

[Notional]

18 minutes ago, Trixanity said:

I think he's referencing the AMD initiative to convert CUDA code to openCL. So it's kinda a yes and no to that. AMD cards will never support CUDA, however if CUDA applications can easily be ported to openCL, they'll have a reasonably similar outcome as if they were natively supporting CUDA (I guess).

Yeah, but that requires the developers to actively port their CUDA programming to OpenCL/C language. AMD doesn't support CUDA and they don't have a license.

[Citadelen]

41 minutes ago, GidonsClaw said:

Conjecture, any backing to that?

It was an informed guess, we know AMD is focusing on low power and efficiency with Polaris, and we know Nvidia will be focusing on compute, which will remove the efficiency advantage they have with Maxwell.

[Jahramika]

4 hours ago, Notional said:

What are you talking about? There is less than 0% chance that Polaris will have CUDA support.

Not support but code compatibility. Well ya support CUDA can be ported open CL

[DocSwag]

5 hours ago, Jahramika said:

Yes

It depends. Polaris will still be on GCN and therefore can't directly run CUDA, but AMD does have that initiative to make a CUDA to GCN compiler (I'm assuming to change CUDA code so it can run on gcn). It depends if they have it out by Polaris.

[stconquest]

1 hour ago, DocSwag said:

It depends. Polaris will still be on GCN and therefore can't directly run CUDA, but AMD does have that initiative to make a CUDA to GCN compiler (I'm assuming to change CUDA code so it can run on gcn). It depends if they have it out by Polaris.

http://www.overclock3d.net/articles/gpu_displays/amd_gets_cuda_compilers/1

 

I was simply wondering if Polaris is to be the first implementation.  I know it is coming sooner or later.

[Notional]

1 hour ago, Jahramika said:

Not support but code compatibility. Well ya support CUDA can be ported open CL

 

4 minutes ago, stconquest said:

http://www.overclock3d.net/articles/gpu_displays/amd_gets_cuda_compilers/1

 

I was simply wondering if Polaris was to be the first implementation.  I know it is coming sooner or later.

What you are thinking about is AMD's Boltzmann initiative. It consists of a compiler specifically designed to handle GCN and HSA. Furthermore it includes the HIP tool, that can convert most CUDA code automatically into C++, with tools to help developers to manually translate/port the last code themselves.

 

The idea is to make software utilize AMD's technologies better, and to help developers more their software from CUDA to OpenCL/C++. In other words, there is 0 CUDA support in GCN, and no license for AMD either.

 

http://www.amd.com/en-us/press-releases/Pages/boltzmann-initiative-2015nov16.aspx

[patrickjp93]

14 minutes ago, Notional said:

 

What you are thinking about is AMD's Boltzmann initiative. It consists of a compiler specifically designed to handle GCN and HSA. Furthermore it includes the HIP tool, that can convert most CUDA code automatically into C++, with tools to help developers to manually translate/port the last code themselves.

 

The idea is to make software utilize AMD's technologies better, and to help developers more their software from CUDA to OpenCL/C++. In other words, there is 0 CUDA support in GCN, and no license for AMD either.

 

http://www.amd.com/en-us/press-releases/Pages/boltzmann-initiative-2015nov16.aspx

It'll be interesting to see which standard wins out: OpenMP/OpenACC or OpenCL for graphics acceleration. 

[Notional]

20 minutes ago, patrickjp93 said:

It'll be interesting to see which standard wins out: OpenMP/OpenACC or OpenCL for graphics acceleration. 

Maybe it's more relevant to ask if they can coexist. If it's easy to port code from one standard to another, if would be better to make several versions of a piece of software optimized for each piece of hardware. There is no way in hell NVidia is giving up one of their proprietary IP's unless forced to. Having NVidia dictate what to use, has proven to be very bad for the industry as a whole (see GameWorks, performance wise and Gsync price wise).

[patrickjp93]

3 minutes ago, Notional said:

Maybe it's more relevant to ask if they can coexist. If it's easy to port code from one standard to another, if would be better to make several versions of a piece of software optimized for each piece of hardware. There is no way in hell NVidia is giving up one of their proprietary IP's unless forced to. Having NVidia dictate what to use, has proven to be very bad for the industry as a whole (see GameWorks, performance wise and Gsync price wise).

OpenCL is not so nice about it. I mean, if you have a cross-compiler, then sure, but there will always be a contest between integration and technological expansion. C++ is still around integrating new programming paradigms and concepts. OpenMP absorbed heterogeneous acceleration from OpenCL, but at the same time programming in OpenMP is considerably easier.

[dalekphalm]

22 hours ago, Jahramika said:

Yes

See below - While you understand that it's a code compiler, it's still misleading to say yes, when he specifically asked about CUDA support.

21 hours ago, Notional said:

What are you talking about? There is less than 0% chance that Polaris will have CUDA support.

Agreed - people were totally misinterpreting that particular topic, saying AMD had a CUDA license finally, and that we'd see native CUDA execution support eventually on GPU's - all of which were either straight out wrong, or simply conjecture.

21 hours ago, Trixanity said:

I think he's referencing the AMD initiative to convert CUDA code to openCL. So it's kinda a yes and no to that. AMD cards will never support CUDA, however if CUDA applications can easily be ported to openCL, they'll have a reasonably similar outcome as if they were natively supporting CUDA (I guess).

 

21 hours ago, Notional said:

Yeah, but that requires the developers to actively port their CUDA programming to OpenCL/C language. AMD doesn't support CUDA and they don't have a license.

 

17 hours ago, Jahramika said:

Not support but code compatibility. Well ya support CUDA can be ported open CL

Big difference. AMD released a tool to translate raw CUDA code into C++ code - that's not even close to having "CUDA Support".

On 2016-02-14 at 8:38 AM, stconquest said:

Is Polaris introducing CUDA support on AMD?

Short answer: No. Unless AMD acquires a CUDA license, which we have seen no indications that either party (AMD nor NVIDIA) want this. There was a rumour years ago that NVIDIA offered AMD a CUDA license but AMD turned it down, but honestly, I haven't ever seen confirmation or evidence that it was anything other than a rumour.

 

For the time being, if you want to be able to use CUDA enabled features (Eg: Many of the GamesWorks features, or GPU assisted PhysX), you're still forced to buy an NVIDIA GPU.