AMD: Speculation, Rumors, and Discussion (Archive)

[jacozz]

I don't know much about Vega except it will be using HBM2. It could be a "big" die, but could also be 2 dies on a interposer. The reason I suggest this is that Raja told Ryan at PCper, in an interview a couple of month ago that AMD won't do big dies anymore, because of economics and time to market. Now... If you going to make use of all that bandwith HBM2 offers, it makes no sense to use it on a medium sized chip. So, either a semi-big chip at maybe 350-400 mm2, or two medium chips paired together. I really hope for the first alternative.

 

[Dangerman]

I don't know much about Vega except it will be using HBM2. It could be a "big" die, but could also be 2 dies on a interposer. The reason I suggest this is that Raja told Ryan at PCper, in an interview a couple of month ago that AMD won't do big dies anymore, because of economics and time to market. Now... If you going to make use of all that bandwith HBM2 offers, it makes no sense to use it on a medium sized chip. So, either a semi-big chip at maybe 350-400 mm2, or two medium chips paired together. I really hope for the first alternative.

When they mean big die do they mean 596mm2 size or even close to GP102 size? Would hope Vega 11 exists as a close to GP102 competitor. Eyeing for a single GPU for 3440x1440.

 

[xEx]

Do we have new info about Rx 470? they just dissapear from internet lol. Im thinking AMD recall all boards to address the "powergate" before they hit the market.

 

[rbraden]

We know that Vega 10 (GP104 competitor) is Q1 2017 but what about Vega 11/Big Vega? Any rumors pointing out size, FLOPs etc? I feel for gaming Vega 11/Big Vega will do better at higher resolutions & DX12 than GP102.

I am also looking forward to specs being released about either vega, being I need to replace my 7970 in one of my systems. I have tried to have both(yes once their were many) vendors thru each gen since the rendition and voodoo 1/matrox days.
I am a bit behind on the AMD side, and no reason to upgrade the 980ti yet. Looking forward to vega. All good things in all good time.

 

[xEx]

Well Vega 10 were "finish" Typed out? like a month ago so I think leaks will be around November.

 

[ImSpartacus]

I don't know much about Vega except it will be using HBM2. It could be a "big" die, but could also be 2 dies on a interposer. The reason I suggest this is that Raja told Ryan at PCper, in an interview a couple of month ago that AMD won't do big dies anymore, because of economics and time to market. Now... If you going to make use of all that bandwith HBM2 offers, it makes no sense to use it on a medium sized chip. So, either a semi-big chip at maybe 350-400 mm2, or two medium chips paired together. I really hope for the first alternative.

I had assumed Vega would be a "last hurrah" for big dies at AMD and Navi would implement some kind of multi-chip solution.

But who knows, maybe big Vega is two Polaris 10-like die and little Vega is two Polaris 11-like die.

 

[rbraden]

I had assumed Vega would be a "last hurrah" for big dies at AMD and Navi would implement some kind of multi-chip solution.

But who knows, maybe big Vega is two Polaris 10-like die and little Vega is two Polaris 11-like die.

I thought I read somewhere that vega would be 4096 shaders, though that could have been a drunken dream I had.

 

[xEx]

Well Raja does said that for him multi GPU is the future and is better in every way(if you can take full advantage over it)

first rumors said vega was 4096 cores indeed.

 

[rbraden]

Well Raja does said that for him multi GPU is the future and is better in every way(if you can take full advantage over it)

first rumors said vega was 4096 cores indeed.

Please excuse my ignorance, but I recall AMD buying a company awhile back, something to do with coherent fabric. Would this only be useful for cpu's or is this something that could be used for gpu's also.
I know nothing about coherent fabric other than skimming over it a few years back.

And sorry for getting off topic.

 

[3dilettante]

Please excuse my ignorance, but I recall AMD buying a company awhile back, something to do with coherent fabric. Would this only be useful for cpu's or is this something that could be used for gpu's also.
I know nothing about coherent fabric other than skimming over it a few years back.

And sorry for getting off topic.

It bought SeaMicro, which created a server fabric for setting up high-density server products.
It was not coherent, and the SeaMicro business was abandoned.
AMD noted it could use the IP in the future, but no examples of the fabric being used outside of the ASIC solution that predated purchase have come to pass. Pretty much the whole purchase price was written off as a loss, so it seems like AMD did not see much worth in what was left.

 

[rbraden]

It bought SeaMicro, which created a server fabric for setting up high-density server products.
It was not coherent, and the SeaMicro business was abandoned.
AMD noted it could use the IP in the future, but no examples of the fabric being used outside of the ASIC solution that predated purchase have come to pass. Pretty much the whole purchase price was written off as a loss, so it seems like AMD did not see much worth in what was left.

Thank you. I just remember reading about the purchase a bit back.
The more you learn.

 

[ImSpartacus]

I thought I read somewhere that vega would be 4096 shaders, though that could have been a drunken dream I had.

Yes, the classic rumor from an AMD engineer's LinkedIn page (now deleted) claimed that Greenland had 4096 shaders.

Personally, I'm still partial to that rumor and I think that AMD wouldn't try to implement a multi-chip solution so soon after the new process (AMD is likely short on R&D capability at the moment).

 

[rbraden]

Yes, the classic rumor from an AMD engineer's LinkedIn page (now deleted) claimed that Greenland had 4096 shaders.

Personally, I'm still partial to that rumor and I think that AMD wouldn't try to implement a multi-chip solution so soon after the new process (AMD is likely short on R&D capability at the moment).

Does anyone happen to know what Graphic IP 9 entails? Would this be GCN 1.5 or something else entirely? I am more than likely totally off the mark here. Just trying to wrap my head around some of this.

 

[3dilettante]

But who knows, maybe big Vega is two Polaris 10-like die and little Vega is two Polaris 11-like die.

Wouldn't that just make little Vega run right into Polaris?

Does anyone happen to know what Graphic IP 9 entails? Would this be GCN 1.5 or something else entirely? I am more than likely totally off the mark here. Just trying to wrap my head around some of this.

AMD has not disclosed much about Vega. Those that would know would not be free to discuss it, which is one probable reason why the LinkedIn entry was deleted.
The trivial answer is that IP 9 entails it is different from 8. AMD has actually changed its criteria for when it bumps up numbers with the latest generation, but it may sufficiently different from what went into 8 that the difference is visible to drivers or lower-level software. There are some things in the full DX12 feature set AMD still doesn't support, so those might be nice to provide. GCN might benefit from some kind of architectural change.

What a number AMD arbitrarily decides to put on a thing like GCN 1, 1.1, 1.5, 2, 3, 4, etc. means is whatever AMD feels like assigning. It is helpful in trying to get something meaningful out of AMD's PR if there's some method to the madness, but AMD could not be bothered for years.
The now more regular GCN 1,2,3,4 numbering is evidence that it has noticed that it did not help to make everyone struggle to figure out what AMD meant by anything.

 

[ImSpartacus]

Wouldn't that just make little Vega run right into Polaris?

Lol, yes. For a moment there, I think I thought that Polaris 11 had like 25 CUs. Yeah, no.

 

[xEx]

Interesting reading and nice site.

http://gpuopen.com/texel-shading/

 

[xEx]

Btw do we know the lunch date for the Nitro?

 

[pharma]

RX 480 Sapphire Nitro review ...
http://www.eteknix.com/sapphire-nitro-rx-480-oc-8gb-graphics-card-review/

 

[xEx]

the 4GB for 220 is the best cost/benefits right now.

 

[Kaotik]

We know that Vega 10 (GP104 competitor) is Q1 2017 but what about Vega 11/Big Vega? Any rumors pointing out size, FLOPs etc? I feel for gaming Vega 11/Big Vega will do better at higher resolutions & DX12 than GP102.

All we have currently is roadmap with Vega at 2016/2017 newyear (which itself means nothing, AMD has previously marked GPUs for the next year of when they're actually released), and rumours of Vega being pushed early to October