AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Anarchist4000]

Im really not quite sure about the " 2018-2019" roadmap. Why keep Vega 10 if you have Vega20 on 7nm. ( who seems only add DP and GMI ) ? Ofc without having the context of thoses internal slides ( if really sorted by AMD ) its a bit hard to know if the roadmap is really accurate.

That's also getting into Navi timelines, which may not have the FP64 working initially. Similar timeline as Vega with graphics parts first and professional last. Another interesting consideration might be those Intel rumors affecting the design.

That would seemingly make a discrete Vega with 1/2 rate DPFP easier, rather than pushing the specs of the supposed lead chip whose name showed up early in driver strings to another node.

The more I think about it the more sense it makes. They've already got a SSD, network controller, USB-C for displayout(?) and who knows what else on the board. Might as well go all in with the CPU. Easy way to get a bunch of Zen cores into a rack as well.

 

[Blazkowicz]

Vega 20 looks like a GPU for "big iron" like GP100 is (or GK210 in a more limited sense, before that)

Might also be late 2018 for Vega 20, early 2019 for Navi 10 for a very liberal definition of "late".
Vega 20 may be more better than Vega 10, but will cost a ton and if you're after a consumer GPU, then Navi 10 will likely be better anyway. Such as having display outputs for example, perhaps more CUs.

 

[ImSpartacus]

Vega 20 looks like a GPU for "big iron" like GP100 is (or GK210 in a more limited sense, before that)

Might also be late 2018 for Vega 20, early 2019 for Navi 10 for a very liberal definition of "late".
Vega 20 may be more better than Vega 10, but will cost a ton and if you're after a consumer GPU, then Navi 10 will likely be better anyway. Such as having display outputs for example, perhaps more CUs.

Do you think having only 8GB of hbm will hurt Vega in competitions with gp100?

I mean, gp100 is a pretty beefy gpu.

 

[lanek]

Do you think having only 8GB of hbm will hurt Vega in competitions with gp100?

I mean, gp100 is a pretty beefy gpu.

Vega 10 have allways been shown ( including during the Vega preview at CES ) with 8 and 16GB HBM2 configuration. But im pretty sure the MI25 have 16GB HBM2. ( 2 stacks with 8GB on each stack ).

Its the version who have been demoed ( the gaming part obviously ) who should have 8GB, and that dont mean that a 16Gb could not come later. )

If the question is too know in the gaming space ? well the TitanXP have 12GB DDR5 but it is GP102, and the 1080TI will surellly have 6 or 12GB configuration... ( if we dont have a surprise from Nvidia who release another gpu based on GP100)

 

[pTmdfx]

None of the Vega products quite match features attributed to Greenland via Linkedin or the HPC APU rumors. The supposed lead chip has 1/2 rate DPFP, with the "APU" using it having PCIe 3.0. I suppose there could be a wrinkle related to the CPU component that could explain that, although why AMD would hold back on double-precision for compute products if it were there is unclear.

Perhaps they predict that grabbing shares from CUDA dominated DPFP intensive spaces would be harder than the growing segments in machine learning, hmm? So they rather to save time by not enabling or validating half rate DPFP in Vega 10. But still, given that even Bistrol Ridge got half rate DPFP, it seems a rather awkward move if true.

 

[pTmdfx]

The more I think about it the more sense it makes. They've already got a SSD, network controller, USB-C for displayout(?) and who knows what else on the board. Might as well go all in with the CPU. Easy way to get a bunch of Zen cores into a rack as well.

The alleged HPC/workstation APU is said to have a GPU spec similar to the alleged Vega 10 though, and Vega 10 is already at 500+. It is hard to imagine how the alleged 16-core CPU SoC can be packed inside, while providing the alleged I/O spec. Moreover, masks are so costly that I honestly don't see why AMD wouldn't reuse the masks as much as possible.

HPC/workstation APU is a niche at this moment with no value-added for its extra cost even when it is reused in consumer/workstation segments, other than being a serious HSA prototype platform which Raven Ridge might do equally well.

It feels more like the workstation APU having been scrapped despite being on the plan of record, since the demand or value-added for tighter integration (contract?) is just not strong enough. Though it is always a chicken-and-egg problem. Another possibility is that... AMD doesn't want a DPFP card to compete with its single-socket HPC/workstation APU, but this can't explain Vega 20.

 

[xEx]

Do you think having only 8GB of hbm will hurt Vega in competitions with gp100?

I mean, gp100 is a pretty beefy gpu.

For games I don't. AMD said that games even at 4k ultra rarely use more than 4gb of ram. They allocate all what they can but only actively use no more than 4gb. That was the reason(as they says) they redesign the memory system to be more efficient on how they manage the memory allocation and only use the ram for work that it's been used actively and the system memory and storage for the rest.

On professional work idk. I think there is the more the better.

Enviado desde mi HTC One mediante Tapatalk

 

[lanek]

Vega 20 looks like a GPU for "big iron" like GP100 is (or GK210 in a more limited sense, before that)

Might also be late 2018 for Vega 20, early 2019 for Navi 10 for a very liberal definition of "late".
Vega 20 may be more better than Vega 10, but will cost a ton and if you're after a consumer GPU, then Navi 10 will likely be better anyway. Such as having display outputs for example, perhaps more CUs.

Yes i have think to that too, specially, we speak about professional gpu's, the timeline could be quite different.

If we look at the initial roadmap provided by AMD a few month ago, Vega is pointed at 2017 (Q1 ), Navi at 2018.. Offcourse they was no mention of Vega20, so roadmap could have slighty change .

 

[lanek]

For games I don't. AMD said that games even at 4k ultra rarely use more than 4gb of ram. They allocate all what they can but only actively use no more than 4gb. That was the reason(as they says) they redesign the memory system to be more efficient on how they manage the memory allocation and only use the ram for work that it's been used actively and the system memory and storage for the rest.

On professional work idk. I think there is the more the better.

Enviado desde mi HTC One mediante Tapatalk

Actually GP100 baed Tesla have too "only" 16GB of HBM2, Same as Vega, but as GP100 have 4Histacks ( 4x4GB) they could upgrade later to 4x8GB. Vega10 (2Histack) will not been able to do it, untill Vega 20 (4Histacks)..

Now, as Vega push a bit the question of storage possibility (Nvram etc ), that change a bit the equation, as Nvidia will offer this with Volta only.

 

[xEx]

Actually GP100 baed Tesla have too "only" 16GB of HBM2, Same as Vega, but as GP100 have 4Histacks ( 4x4GB) they could upgrade later to 4x8GB. Vega10 (2Histack) will not been able to do it, untill Vega 20 (4Histacks).

Yes but amd hace 'the cube' and make 4 vga together and have the ssg and put ssds on them.

I mean both have 'features' over the other.

BTW does the ssg thing is useful for anything else of video editing?

Enviado desde mi HTC One mediante Tapatalk

 

[3dilettante]

Perhaps they predict that grabbing shares from CUDA dominated DPFP intensive spaces would be harder than the growing segments in machine learning, hmm?

It's harder if AMD needs to do so with Hawaii. AMD has more of a history and some presence in that kind of workload, where GPU-based machine learning is a new field where Nvidia has laid down all the groundwork. There's still serious money in that space, although this does see to hint that AMD is rather shut out of some large system buys for at least a product generation or more if all the enhanced connectivity and IO that a big government or science cluster would want is disqualified due to the GPU it is tied to.

So they rather to save time by not enabling or validating half rate DPFP in Vega 10. But still, given that even Bistrol Ridge got half rate DPFP, it seems a rather awkward move if true.

I'm not sure what significant amount of time AMD is saving. The idea behind the configurable DP throughput of GCN is that it is something that is readily adjustable as part of the architecture's foundation. Hawaii's development cycle didn't seem to be overly extended due to its 1/2 rate DP. The big changes that really need heavy validation work shouldn't be an ALU mode.

 

[lanek]

Yes but amd hace 'the cube' and make 4 vga together and have the ssg and put ssds on them.

I mean both have 'features' over the other.

BTW does the ssg thing is useful for anything else of video editing?

Enviado desde mi HTC One mediante Tapatalk

Yes, ofc, VFX, 3D animation, 3D render (CG), specially with the move to scanned textures/object... (include everything who use now 3D modelisation on scientific aspect). And surely a lot of other domain in computing.

 

[xEx]

Yes, ofc, VFX, 3D animation, 3D render (CG), specially with the move to scanned textures/object... (include everything who use now 3D modelisation on scientific aspect). And surely a lot of other domain in computing.

( note, i had allrready edited my post with the "storage possibility who offer Vega )

What about gaming? Could be better let's say use 6gb of ram and 10 of ssd storage or some combination like that? Could it make vga with high storage capabilities cheaper?

Enviado desde mi HTC One mediante Tapatalk

 

[lanek]

What about gaming? Could be better let's say use 6gb of ram and 10 of ssd storage or some combination like that? Could it make vga with high storage capabilities cheaper?

Enviado desde mi HTC One mediante Tapatalk

For gaming, i let developpers respond, but i could imagine that data who dont need to be available immediately could lay on Nvram/ssd storage ( onboard with the gpu's with direct and fast access ) instead of take Vram space as it is mostly done today. But that will surely need a nice rewrite of game engine.

One example could be extremely large texture streaming,

 

[xEx]

For gaming, i let developpers respond, but i could imagine that data who dont need to be available immediately could lay on Nvram/ssd storage ( onboard with the gpu's with direct and fast access ) instead of take Vram space as it is mostly done today.

I think AMD may be leading to this with its new memory management system.

Enviado desde mi HTC One mediante Tapatalk

 

[lanek]

I think AMD may be leading to this with its new memory management system.

Enviado desde mi HTC One mediante Tapatalk

One example could be extremely large texture streaming..

 

[pTmdfx]

I think AMD may be leading to this with its new memory management system.

Nvidia already has hot page migration in CUDA. It doesn't seem to be in the graphics stack though. In the end, if one can migrate any pages in the process virtual address space, streaming memory mapped files should probably be a natural extension.

 

[ImSpartacus]

Vega 10 have allways been shown ( including during the Vega preview at CES ) with 8 and 16GB HBM2 configuration. But im pretty sure the MI25 have 16GB HBM2. ( 2 stacks with 8GB on each stack ).

That's good to hear. Though that would require a larger interposer, right? Surely the pro market would justify such expenses, so that shouldn't be an issue.

Its the version who have been demoed ( the gaming part obviously ) who should have 8GB, and that dont mean that a 16Gb could not come later. )

If the question is too know in the gaming space ? well the TitanXP have 12GB DDR5 but it is GP102, and the 1080TI will surellly have 6 or 12GB configuration... ( if we dont have a surprise from Nvidia who release another gpu based on GP100)

For games I don't. AMD said that games even at 4k ultra rarely use more than 4gb of ram. They allocate all what they can but only actively use no more than 4gb. That was the reason(as they says) they redesign the memory system to be more efficient on how they manage the memory allocation and only use the ram for work that it's been used actively and the system memory and storage for the rest.

On professional work idk. I think there is the more the better.

I was mostly concerned with professional use cases, hence the gp100 mention. I know next to nothing about professional uses for graphics cards, but my understanding is that increased vram capacity makes life easier (easier to fit bigger datasets in memory, etc).

8GB oughta be plenty for a halo gaming card in 2017 (though future 384-bit gp102 cards could sport 12 or 24 GB with existing 8Gb GDDR5X offerings, so that could make the marketing more challenging for AMD).

 

[lanek]

That's good to hear. Though that would require a larger interposer, right? Surely the pro market would justify such expenses, so that shouldn't be an issue.





I was mostly concerned with professional use cases, hence the gp100 mention. I know next to nothing about professional uses for graphics cards, but my understanding is that increased vram capacity makes life easier (easier to fit bigger datasets in memory, etc).

8GB oughta be plenty for a halo gaming card in 2017 (though future 384-bit gp102 cards could sport 12 or 24 GB with existing 8Gb GDDR5X offerings, so that could make the marketing more challenging for AMD).

No no, Vega is allready capable of use 2x 8GB stack, its just that the 8GB stacked ram are not there yet or too much costly for the gaming part at this moment. But it is a different story for professional gpu's.
In fact, the MI25 who have been presented by AMD should allready sport 2x8Gb stack. ( and this card will be shipped before gaming parts if i have understand it well ).

As for GP102, yes, but if we look that TitanXP sport 12GB allready , i doubt the 1080TI will sport more, 12GB is allready completely an overkill on gaming gpu's. 8GB today is allready an overkill.

 

[Blazkowicz]

What about gaming? Could be better let's say use 6gb of ram and 10 of ssd storage or some combination like that? Could it make vga with high storage capabilities cheaper?

Enviado desde mi HTC One mediante Tapatalk

Like, you play some game made like id's Rage but textures are streamed from the NVRAM? Shouldn't be hard, but if the needed hardware is only found on $3K or $5K Fire Pro cards and such, there won't much of a market for the feature.

For less money, you might load a desktop up with 128 GB of rather fast RAM, copy the whole game into a ramdisk and live with the mere PCIe 16x 3.0 interface between your CPU and graphics card.