videocardzAMD’s Greenland/Vega 10 GPU to feature 4096 Stream Processors

[csbin]

http://videocardz.com/58665/amds-greenlandvega-10-gpu-to-feature-4096-stream-processors

Guys over at 3DCenter discovered a very interesting piece of information. According to Yu Zheng LikenIn profile, who is a research and development manager at AMD, company has been working on new flagship graphics processing unit also known as Greenland or Vega 10.
The profile reveals that Greenland is the leading project of Graphic IP v9.0. In other words, this is the new flagship GPU that will succeed the most powerful Fiji processor currently offered by AMD.
The graphic IP level 9.0 most likely stands for Vega architecture (the first to feature HBM2 memory). The project is tied with AMD’s SoC (System on a Chip) v15 architecture, which could suggest that future AMD APUs could also be somewhat connected to Polaris and Vega GPUs.

 

[happy medium]

The graphic IP level 9.0 most likely stands for Vega architecture (the first to feature HBM2 memory).

that settles that.

 

[Silverforce11]

Looks like this next-gen GCN is a huge change. We've been expecting increased SP counts, but it's actually declining comparatively.

For Vega 10, 4096 SP is not an improvement over Fiji on SP count.

The SP count on Polaris 10 is also pretty low for it's die size and known 14 ff density. Basically, 232mm2 on 14ff with 2.2x density (some sources claim as high as 2.4x), is equivalent to Hawaii or greater, but it's got a reduction in SP.

 

[Adored]

Fiji can't saturate 4096 SPs without async so it shouldn't be a surprise that AMD doesn't want to go much further down that route.

 

[msi2]

We don't really know... Could be Vega 11, that's all speculation.

 

[JDG1980]

Looks like this next-gen GCN is a huge change. We've been expecting increased SP counts, but it's actually declining comparatively.

I expect "IPC" (per-clock, per-shader performance) to go up by a measurable margin between GCN 1.2 and Polaris/Vega. However, I think there's a chance that even more of the gains will be the result of clock speed increases. 14nm FinFET should enable considerably higher clocks than 28nm planar.

If you could run Hawaii's core at ~1.4 GHz (and used GDDR5X or HBM2 so that it wouldn't be starved for bandwidth at that speed) it would kick Fiji's ass. And Hawaii is only a GCN 1.1 part.

 

[Hi-Fi Man]

Is it really known whether these GPUs are going to be made with Samsung/GFs 14nm LPP node? If so then NVIDIA might be at a disadvantage because they'll most likely be using TSMCs 20nm FinFET node (16nm FF+).

 

[raghu78]

Looks like this next-gen GCN is a huge change. We've been expecting increased SP counts, but it's actually declining comparatively.

For Vega 10, 4096 SP is not an improvement over Fiji on SP count.

The SP count on Polaris 10 is also pretty low for it's die size and known 14 ff density. Basically, 232mm2 on 14ff with 2.2x density (some sources claim as high as 2.4x), is equivalent to Hawaii or greater, but it's got a reduction in SP.

We know there are going to be two Vega chips. Vega 10 ad Vega 11. So I think its possible we see atleast a doubling of Polaris 10 in terms of sp. I would not be surprised to see Vega 11 at 3840-4096 sp and Vega 10 at 5120sp.

 

[Silverforce11]

We know there are going to be two Vega chips. Vega 10 ad Vega 11. So I think its possible we see atleast a doubling of Polaris 10 in terms of sp. I would not be surprised to see Vega 11 at 3840-4096 sp and Vega 10 at 5120sp.

I agree, as there's going to be some segment issues if Vega 10 is 4096, then Vega 11 has to squeeze in between that and 2560 SP. There's not much room for cut-down SKUs on Vega 10 and 11.

Whereas Vega 11 4096 SP, the cut down can be ~3.6K SP giving it a decent performance edge over Polaris 10.

 

[tenks]

I think its strange that after launching Fiji with HBM, they're skipping that product line with the polaris generation/arch and then bringing it back with Vega.

We don't really know... Could be Vega 11, that's all speculation.

Actually we do know based on simple logic. This is the chip using HBM2..which is for high-end skus only. They all used the nomenclature "a full 4096 sps" which means its maxed out...as in the word 'full'..So logic would tell us there isn't another chip with more SP's and that this is Vega 10, the big chip, project greenland, which has been talked about forever.

They wouldn't say FULL 4096 sp if it could have more and they wouldnt make such a big deal out of project "greenland" if it wasn't the flagship gpu.

Pretty simple.

 

[3DVagabond]

I think its strange that after launching Fiji with HBM, they're skipping that product line with the polaris generation/arch and then bringing it back with Vega.



Actually we do know based on simple logic. This is the chip using HBM2..which is for high-end skus only. They all used the nomenclature "a full 4096 sps" which means its maxed out...as in the word 'full'..So logic would tell us there isn't another chip with more SP's and that this is Vega 10, the big chip, project greenland, which has been talked about forever.

They wouldn't say FULL 4096 sp if it could have more and they wouldnt make such a big deal out of project "greenland" if it wasn't the flagship gpu.

Pretty simple.

It could be the Flagship like Hawaii was. They could still have a bigger chip in the works for 1+ years from now.

 

[MrTeal]

I think its strange that after launching Fiji with HBM, they're skipping that product line with the polaris generation/arch and then bringing it back with Vega.



Actually we do know based on simple logic. This is the chip using HBM2..which is for high-end skus only. They all used the nomenclature "a full 4096 sps" which means its maxed out...as in the word 'full'..So logic would tell us there isn't another chip with more SP's and that this is Vega 10, the big chip, project greenland, which has been talked about forever.

They wouldn't say FULL 4096 sp if it could have more and they wouldnt make such a big deal out of project "greenland" if it wasn't the flagship gpu.

Pretty simple.

Not necessarily, both Vega chips are likely to have HBM2. It's also a LinkedIn profile, so wording is often exaggerated, and it's obviously written by an ESL person. The lead chip of its generation doesn't necessarily mean best, it could just as well mean first. "The full capacity of 4096 shader processors" is just as (or more) likely to mean that the full chip has 4096 shaders than AMD is capable of producing in a GCN GPU.

 

[Head1985]

I think its vega 11.Vega 10 should be 2x polaris 10(or more)so 5120SP+.
This card with 4096SP should compete against GP104.I dont see it compete against GP100.It will be mid-range GPU around 330-350mm2(tahiti replacement).Vega 10 with 5120SP+ should be around 430-450mm2.(hawaii replacement)

Btw 5120SP vs 4096SP is only 25% increase.Its too small gap for new SKU.It should be atleast 30+%.Its wierd.

In my speculations vega 11 have 3840SP and gap was 33% to Flagship vega10.

EDIT:
7870 vs 290x its 1280SP vs 2808SP- hawaii have 119% more SP.Its more than 2x 7870.
Polaris10(7870 successor) vs vega10(hawaii successor) 2560SP vs ???-Vega 10 should have more than 2x more SP than polaris 10.

 

[Krteq]

As I repeatedly told you on another forums, we know nothing about Pascal architecture and chip configurations so these claims are baseless nonsenses.

nVidia didn't show us nothing, there are no driver leaks etc., so how did you manage to get these performance estimates?

 

[ShintaiDK]

As expected, performance /watt is the key and "small" dies due to cost. Raw performance lost.

Hopefully the chip is better balanced than Fiji.

 

[ShintaiDK]

Is it really known whether these GPUs are going to be made with Samsung/GFs 14nm LPP node? If so then NVIDIA might be at a disadvantage because they'll most likely be using TSMCs 20nm FinFET node (16nm FF+).

Glofo 14LPP. TSMC 16FF+ got better electrics. A10 is TSMC 16FF+ as well.

With the WSA AMD dont have a choice like the rest.

 

[Adored]

Glofo 14LPP. TSMC 16FF+ got better electrics. A10 is TSMC 16FF+ as well.

With the WSA AMD dont have a choice like the rest.

So why did they go to Samsung for more capacity in 2017 instead of TSMC?

 

[itsmydamnation]

Glofo 14LPP. TSMC 16FF+ got better electrics.

you claimed that last time and link you provided proved nothing of the sort, at this stage its nothing but FUD.

 

[ShintaiDK]

So why did they go to Samsung for more capacity in 2017 instead of TSMC?

Who? AMD? I doubt they get anything made outside GloFo unless they can get much higher volume than today so the WSA gets filled. And I dont see that happening.

Remember, AMD is locked to Glofo not by will, but by force till may 2nd 2024. They simply dont have a choice unless they can have a volume higher than the WSA requirement. But even then, all the WSA allocation needs to be made at Glofo.

 

[Adored]

Who? AMD? I doubt they get anything made outside GloFo unless they can get much higher volume than today so the WSA gets filled. And I dont see that happening.

Remember, AMD is locked to Glofo not by will, but by force till may 2nd 2024. They simply dont have a choice unless they can have a volume higher than the WSA requirement. But even then, all the WSA allocation needs to be made at Glofo.

What is your estimate on how much AMD spends on wafers every year?

 

[ShintaiDK]

What is your estimate on how much AMD spends on wafers every year?

You can lookup the WSA amount and how much they have to pay when its unfilled.

 

[Gloomy]

As expected, performance /watt is the key and "small" dies due to cost. Raw performance lost.

How can you say that with any certainty?

Maxwell for example already showed that more performance can be extracted from fewer shaders in practice , compared to kepler (which was poorly balanced as fiji/gcn in general, just in the opposite direction).

 

[Adored]

You can lookup the WSA amount and how much they have to pay when its unfilled.

When was the last time they had to pay for unfilled quota at globalfoundries?

 

[rainy]

Maxwell for example already showed that more performance can be extracted from fewer shaders in practice , compared to kepler (which was poorly balanced as fiji/gcn in general, just in the opposite direction).

Just a simple proof of what you saying:
https://www.techpowerup.com/gpudb/2512/geforce-gtx-780-ti.html
https://www.techpowerup.com/gpudb/2621/geforce-gtx-980.html

 

[ShintaiDK]

How can you say that with any certainty?

Maxwell for example already showed that more performance can be extracted from fewer shaders in practice , compared to kepler (which was poorly balanced as fiji/gcn in general, just in the opposite direction).

Remember size as well.

GTX680=294mm2 ~3.5B (32ROP)
GTX980=398mm2 ~5.2B (64ROP)

I fully agree on the unbalanced chip as I already stated. And for that matter every time its not shader limited Vega 10 can look much better, assuming ROP increase. And hopefully they increase it to 128. But I doubt shader power will change much at all and I will guess they will look around equal gflops wise.

I expect Nvidia to go the exact same route as well. Only with the GP100 as the joker due to Knights Landing. If GP100 cant compete with this, Nvidia loses the HPC. But any GP100 chips will cost more than your first born assuming they can. Another thing people have to remember, there are no cost savings transistor wise. All the current leaked chips seems to be more or less shrinked 28nm in terms of size with little to no increase. But again, performance/watt and low power is what the 99% crowd wants. In that relation this forum lives in its own world. So financially they are doing the right move. If I had to guess, Vega 10would end up in the 175-200W area as top bin.