Vega refresh - Expected? How might it look?

[Guru]

AMD can only be redeemed and saved if game developers start using all of Vega's techniques and features, but considering we are still essentially stuck at DX11 and most engines are at least 4-5 years old by this point, I don't see too many developers taking it upon themselves to use additional resources to code for AMD cards, especially as they are not the market leader for gamers.

We see when a game is built to take full advantage of Vega that the Vega 64 beats the 1080ti, Wolfenstein 2 shows how powerful Vega can be, but the issue is the design is too specific, too DX12 that any older game, older engine, older code, etc... will have the Vega's struggling. That is why we see some independant games released in Unity or Unreal engines have Vega 64 on par with the 1070.

 

[DisEnchantment]

Videocardz have been spot on with the Vega leaks so far.
....

Vega 10 x2
Thus far it would seem that the Vega 10 x2 is dead.

Vega 11
That said, Vega 11 is still a mystery to me, is it really RR with 11 CU or really an architecture.
References to EEC for Vega 10, 11, 12

https://videocardz.com/71280/amd-vega-10-vega-11-vega-12-and-vega-20-confirmed-by-eec

Then here James Prior saying Vega 11 is RR

https://videocardz.com/74260/amds-james-prior-talks-ryzen-2-and-vega-11

But I leave it at that.

Vega 12
During AMD Tech Day, Vega Mobile was announced and we also learnt the following.

However new driver PRs were issued for Kernel 4.17 which would seem to me strange that AMD would not wait for 4.18 anymore.
This means launch Window is fairly close.
This new architecture is Vega 12, and it is not the Vega Mobile or the Vega for Kaby Lake G
I could be wrong, but according to bridgman [AMD SW guy], the product has not been announced yet.

https://www.phoronix.com/forums/for...ver-patches-for-vega-12?p=1014511#post1014511

https://www.phoronix.com/forums/for...-into-linux-4-17-kernel?p=1014971#post1014971

I hope bridgman is not rebuked for saying this. I admire the guy.

AMD is strangely quiet about this, they mentioned Vega at 12 nm as well but remained quiet afterwards, they did not confirm or deny.
I checked the code for AMD kernel mode driver and Vega 12 is much closer to Vega 10 than RR (15DD)
Memory wise it looks like it will be HBM2 like Vega 10. I wonder if it will use newer HBMs unlike the last year's 1.6 Gbps ones, Vega is BW starved it looks to me.
It is still GFX9 architecturally, at least what I could infer from the code.

Commits are in this branch

https://cgit.freedesktop.org/~agd5f/linux/log/?h=drm-next-4.17

This Vega nomenclature is such a confusion.
Vega 12nm, Vega 12, Vega 3/6/8/11 CU, Vega 10, Vega 11?

Vega Line up 10 months ago registered with the EEC.

 

[jpiniero]

Could just be a desktop version of Vega Mobile, aimed at replacing the 570 and 580 (although the MSRP would obviously be much higher).

 

[Glo.]

Vega 12 is the GPU that has been announced as Vega Mobile.

 

[DisEnchantment]

Vega 12 is the GPU that has been announced as Vega Mobile.

bridgman said it was not yet announced, do you have some sources? He is a long time AMD SW guy and quite renowned in mesa and drm circles.
But I could have interpreted some things incorrectly.

 

[Glo.]

bridgman said it was not yet announced, do you have some sources? He is a long time AMD SW guy and quite renowned in mesa and drm circles.
But I could have interpreted some things incorrectly.

PRODUCT was not announced.

The chip was. Vega Mobile. Product is Vega 12, and all of SKUs it will spawn.

 

[zinfamous]

Was Vega 10 x2 ever really alive? I recall that only ever existing in the hearts and minds of speculating tech journalists and consumers. If AMD actually posted something like that in a product slide, I don't remember it.

 

[richaron]

Was Vega 10 x2 ever really alive? I recall that only ever existing in the hearts and minds of speculating tech journalists and consumers. If AMD actually posted something like that in a product slide, I don't remember it.

As far as I can tell most of the effort put into Vega was on modularity/connectivity (IF & cache controllers building upon HSA/hUMA efforts) plus introducing new hardware instruction paths. I suspect if they put the same amount of effort simply into optimizations there would be a very different view of the product. I kinda see Vega as a stepping stone in long term goals and almost a sacrificial lamb.

That being said I'm sure an x2 was on the cards (...), it should have played into the modularity ideal and it's strengths. But with the realization of real world performance, power consumption, and lack of software optimizations, I guess the idea was scrapped.

 

[happy medium]

I firmly believe that dual gpu cards died with the $1500 AMD 295 x2/ $999 gtx690, along with SLI/crossfire.
AMD and Nvidia made a new ultra high end segment with larger single chips with cards like Fury X and the Titan /TI series.

 

[EXCellR8]

Dual GPU Vega card would require way too much juice, not to mention it would run incredibly hot on a single PCB without aftermarket or liquid cooling.

i regret not picking up the 295x2 when people were looking to pawn them off cheap (less than $500) but I don't really have much use for it anyway; it's not HBM. I never considered any of the Fury cards because of the cost and by the time Vega rolled out, before the massive cluster of pricing nonsense, I was ready to upgrade from Polaris.

Dual GPU Polaris would actually be kind of neat... did Powercolor ever do one of those?

 

[DisEnchantment]

I find it a little bit funny that we get so much news about a mostly unknown GPU. I'm not complaining, it's just weird.

It's just one more piece of the overall plan falling into place. With a bit of luck this should become the new normal...

https://www.phoronix.com/forums/for...drm-next-for-linux-4-17?p=1015413#post1015413

694X : Kaby Lake G
69AX : Vega 12

This device is already supported in macOS High Sierra
https://egpu.io/forums/mac-setup/ma...nd-radeon-rx-vega-round-3/paged/2/#post-29302

PRODUCT was not announced.

The chip was. Vega Mobile. Product is Vega 12, and all of SKUs it will spawn.

In the EEC listing Vega 12 (AMD FIREPRO VEGA12XT WORKSTATION FCST) is listed as a Workstation part.

Vega 12 might just as well be the mythical Vega 10 x2. 10x2 = 12 Makes perfect sense. . AMD trolling us with the Vega codenames.

 

[Jackie60]

I suspect Vega refresh will be a 25mhz clockspeed improvement with little to show in real world situations. Maybe in 3 years AMD will come good again but I have zero faith in their being competitive this or next year. ‘Poor Volta’ ring any bells? I want to be wrong but I’ve been following this game for 20 years and its part of a familiar cycle.

 

[happy medium]

Vega "Refresh" is going to be a 7nm die shrink of 14nm Vega in 2019.

 

[Stuka87]

Dual GPU Vega card would require way too much juice, not to mention it would run incredibly hot on a single PCB without aftermarket or liquid cooling.

i regret not picking up the 295x2 when people were looking to pawn them off cheap (less than $500) but I don't really have much use for it anyway; it's not HBM. I never considered any of the Fury cards because of the cost and by the time Vega rolled out, before the massive cluster of pricing nonsense, I was ready to upgrade from Polaris.

Dual GPU Polaris would actually be kind of neat... did Powercolor ever do one of those?

Declock the Vega by 10% and it just sips power. Power consumption is only high because AMD chose raw performance over performance-per-watt. A dual card could easily be declocked that much and would still offer a big jump in performance for games that actually support CF.

 

[Krteq]

Vega 10x2 is supposed to be in INSTINCT lineup only.

 

[iBoMbY]

In the EEC listing Vega 12 (AMD FIREPRO VEGA12XT WORKSTATION FCST) is listed as a Workstation part.

Well, exactly what AMD says about the already announced dGPU "Vega Mobile":

Also planned for "Ultrathin workstations" which probably means a 1-slot PCI AIB ...

 

[EXCellR8]

Declock the Vega by 10% and it just sips power. Power consumption is only high because AMD chose raw performance over performance-per-watt. A dual card could easily be declocked that much and would still offer a big jump in performance for games that actually support CF.

Sounds like the die is going to be shrunk down anyway, so they might not need to tame it that much after all. I'd be torn on spending the extra for a dual GPU Vega card with HBM2 since there are so few games that support it well (unless I'm misinformed at the time of this writing) and I can't see availability being very good for the duration of the Vega line.

 

[Krteq]

That "Premium Mobile" section is about Vega in Caby-G (1.7mm height) not "Vega 12"

 

[PeterScott]

That "Premium Mobile" section is about Vega in Caby-G (1.7mm height) not "Vega 12"

It doesn't make any sense that AMD would only have a mobile Vega for Intel, and skip something to support their own CPUs.

Plus I remember seeing some version turn up in one of those online benchmarks that had a more CUs (28?) than Kaby-G.

So I think Vega Mobile, is a product that will soon be released, that is separate from Kaby-G.

 

[Krteq]

It doesn't make any sense that AMD would only have a mobile Vega for Intel, and skip something to support their own CPUs.

Plus I remember seeing some version turn up in one of those online benchmarks that had a more CUs (28?) than Kaby-G.

So I think Vega Mobile, is a product that will soon be released, that is separate from Kaby-G.

They already released Ryzens G in February, do you remember?

 

[PeterScott]

They already released Ryzens G in February, do you remember?

Raven Ridge (2400G) 11 CU parts, are not competitive with Kaby-G 24 CU parts.

I expect AMD will deliver a 24+ CU mobile part, either paired with Ryzen, or on it's own that can be used with Ryzen.

It makes no sense at all for AMD to force anyone who wants a better AMD Vega mobile GPU, to buy their competitors CPU in order to get it.

 

[Krteq]

Why it should be mobile part?

 

[iBoMbY]

That "Premium Mobile" section is about Vega in Caby-G (1.7mm height) not "Vega 12"

Lol, no?

 

[Krteq]

Well... I missed that one

Thx

 

[iBoMbY]

At least this Vega12 doesn't look like a Kaby-G, more like a slower dGPU, to me: