Vega12 begins to show up in AMD's Linux drivers

[Humbug]

AMD devs this week have begun posting updates to their Linux graphics drivers in order to add support for a new discrete graphics architecture code-named 'vega12'. This is not a product name like Vega56 or Vega64, it is an internal code-name similar to vega10 for the already released Vega desktop parts. So the final product variants will be named something else by the marketing team when they are actually released. This is merely a sign that they are close.

 

In the new driver code AMD has reserved 5 PCI IDs for Vega12. This indicates that AMD anticipates an absolute maximum of 5 different SKUs being released based on the Vega12 architecture. Probably different variants tuned to different TDPs, clockspeeds or with different numbers of CUs.

 

Based on general consensus there are two possibilities of what this architecture may be.

option 1- A discrete graphics chip for laptops in order to compete with the Nvidia GTX 1060m

option 2- A new desktop midrange graphics card in order to replace existing Polaris parts such as the RX 580

 

For those of you who are not aware; AMD on Linux supports most of their users via open source graphics drivers. This means that the general public can look at the code and get hints about upcoming products such as this. We still don't know exactly what it is though.

 

Thanks to Michael @ Phoronix for the story.

 

Quote

It's been a while since last hearing anything about the rumored "Vega 12" GPU but coming out this morning are a set of 42 patches providing support for this unreleased GPU within the mainline Linux kernel. Alex Deucher of AMD's Linux driver team sent out the 42 patches this morning providing initial support for Vega 12 within the AMDGPU DRM kernel driver. In total the patches are another sixty thousand lines of kernel code, but the vast majority of that sizable code chunk is auto-generated header files for the GPU registers, etc. The actual new code additions to AMDGPU DRM is relatively small and is mostly re-using existing Vega 10 and Raven Ridge code paths. Digging through the code there aren't any really exciting details revealed. About the only thing worth mentioning is that this Vega 12 enablement has five PCI IDs added, but that's not necessarily how many different SKUs will be released. Often times the GPU vendors reserve extra PCI IDs for possible future use, engineering variants, etc. Those GPU PCI IDs for Vega 12 are listed as 0x69A0, 0x69A1, 0x69A2, 0x69A3, and 0x69AF.

https://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-Vega-12-Patches

 

Quote

Details are still scarce on the "Vega 12" GPU but is to be some new desktop GPU model and most of the speculation seems to be on it being a successor to the Radeon RX 500 "Polaris" series. An AMD representative already confirmed in our forums yesterday that Vega 12 is not about the Vega GPU found on select Intel CPUs. But for now there isn't much information to pass along and these Linux driver patches do not really reveal any useful information and is mostly leveraging existing Vega/Raven code-paths. The RadeonSI Gallium3D patch for adding Vega 12 support is just around 40 lines of code for adding in the Vega 12 name, the five PCI IDs mentioned yesterday, and other very basic alterations. The AMDGPU kernel-side work is what's much more notable. Also out today is a patch with an updated ADDRLIB for Mesa with Vega12 support. ADDRLIB is used for texture addressing and alignment calculations.

https://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-Vega-12

 

Quote

 

AMD developers have already submitted a few rounds of feature work to DRM-Next for Linux 4.17, including enabling DC for all supported GPUs while now they have sent in a last-minute pull request in aiming to get their newly-published "Vega 12" GPU support into the Linux 4.17 kernel. Alex Deucher of AMD sent in this last feature pull to DRM-Next for in turn targeting the Linux 4.17 merge window. There are a few bug fixes and clean-ups for PowerPlay, but most notable is the Vega 12 addition.

 

This Vega 12 addition to the AMDGPU DRM driver isn't a trivial addition but this latest pull request amounts to more than sixty thousand lines of code... Fortunately, a vast majority of that is from auto-generated header files for the GPU's registers, etc. So now there's this pull request that would add the initial Vega 12 support to Linux 4.17. Given that AMD is trying to get it in at the last minute for DRM-Next, it does indicate that the Vega 12 GPU launch could possibly be sooner rather than later with not just waiting around until Linux 4.18. The Linux 4.17 kernel should debut as stable by the middle of June

 

https://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-Vega-12-For-Linux-4.17

 

 

[Mihle]

I am guessing Vega 12 stands for Vega 12 nm. 

[Valentyn]

1 minute ago, Mihle said:

I am guessing Vega 12 stands for Vega 12 nm. 

Vega 64, Vega 56, Vega 12. Could be their baby card for OEMs and power efficiency like the 8 CU in APUs and Intel CPU packages.

[Jurrunio]

Add another option: larger Vega to compete with next gen Nvidia cards.

 

Actually, could it be Vega11 (not the Vega 11 in 2400G) with GDDR instead of HBM? Vega11 leaked in the past was expected to be Polaris' replacement, yet it didnt come which is understandable. HBM is expensive and doesnt make sense when AMD's tactic was to make a cheaper card than Nvidia and sell them in large numbers.

 

Sure, Gamers Nexus did a detailed video, saying  the memory type swap will need a 'memory controller, power delivery, PCB layout, package' redesign, and also hurts Vega performance because it's 'starved of bandwidth'. However, 7 months have passed, even GDDR memory supply is starting to get short, and HBM which is still in the state of 'limited production'. Could AMD decide to cut their losses and do anything they can to make Vega profitable? They clearly stated that Vega GPU production isnt an issue, HBM supply is.

[Humbug]

Just now, Valentyn said:

Vega 64, Vega 56, Vega 12. Could be their baby card for OEMs and power efficiency like the 8 CU in APUs and Intel CPU packages.

Actually; in the driver code Vega56 and Vega64 are referred to as Vega10.

So 12 is not the number of CUs.

 

[Jurrunio]

1 minute ago, Valentyn said:

Vega 64, Vega 56, Vega 12. Could be their baby card for OEMs and power efficiency like the 8 CU in APUs and Intel CPU packages.

Vega 12 here is an internal code name. The year old Vega 10 eventually turned into Vega 64 and 56, the marketing names. this can lead to confusion as AMD has two Vega 11, so it's crucial to distinguish between the two.

[Valentyn]

14 minutes ago, Humbug said:

Actually; in the driver code Vega56 and Vega64 are referred to as Vega10.

So 12 is not the number of CUs.

 

It's still baby Vega then? Like Polaris 10 vs 11 likely. 

I'd assume next Vega would be Vega 20

[Humbug]

7 minutes ago, Jurrunio said:

Add another option: larger Vega to compete with next gen Nvidia cards.

Is there any such thing? Most people seem to think that Vega64 will have to remain the flagship until Navi.

 

Maybe there will be a die shrunk version of Vega64 at some point which allows it to become more efficient and clock higher. I am speculating here but if that does happens at most I expect it to match the 1080ti, I don't expect it to compete with next gen Nvidia flagship.

[Jurrunio]

40 minutes ago, Humbug said:

Is there any such thing? Most people seem to think that Vega64 will have to remain the flagship until Navi.

 

Maybe there will be a die shrunk version of Vega64 at some point which allows it to become more efficient and clock higher. I am speculating here but if that does happens at most I expect it to match the 1080ti, I don't expect it to compete with next gen Nvidia flagship.

Nvidia has a track record of improving things in a linear manner in the past generations starting from the 700 series. The x70 card of the next generation will match the previous generation's x80 Ti. This has been the case when comparing the 780ti to 970 and 980ti to 1070 (even 680 to 770). 

 

At this point, the Vega 64 Liquid only managed to match the GTX 1080, which means the next gen 2070 could have easily beaten it. Yet, Vega 64 LC is already AMD's trump card atm. 7nm process (AMD did say 7nm Vega will come within this year at the start of the year, so probably 31st Dec) could allow potential clock speed increase, but it comes to how TSMC or whoever made the GPUs for AMD performs (see how Ryzen got stuck at 4GHz because TSMC isnt that good at making 14nm FinFet transistors?). Putting their hopes on someone else is clearly not what they want, so making an even larger core to compete with the future 2070 seems logical (and yes, still using its own flagship against Nvidia's cut down high end card).

[Trixanity]

Isn't it just Vega mobile?

The one that's already announced with 28 CUs. They could just take that GPU and put it on a PCI-E card and call it a day.

 

1 hour ago, Jurrunio said:

Nvidia has a track record of improving things in a linear manner in the past generations starting from the 700 series. The x70 card of the next generation will match the previous generation's x80 Ti. This has been the case when comparing the 780ti to 970 and 980ti to 1070 (even 680 to 770). 

 

At this point, the Vega 64 Liquid only managed to match the GTX 1080, which means the next gen 2070 could have easily beaten it. Yet, Vega 64 LC is already AMD's trump card atm. 7nm process (AMD did say 7nm Vega will come within this year at the start of the year, so probably 31st Dec) could allow potential clock speed increase, but it comes to how TSMC or whoever made the GPUs for AMD performs (see how Ryzen got stuck at 4GHz because TSMC isnt that good at making 14nm FinFet transistors?). Putting their hopes on someone else is clearly not what they want, so making an even larger core to compete with the future 2070 seems logical (and yes, still using its own flagship against Nvidia's cut down high end card).

TSMC hasn't made any AMD product that's been released over the last few years. 7nm Vega won't see a consumer release.

 

It's highly unlikely for AMD to release a bigger Vega design. The current Vega 10 chip is already as big as GCN allows and it's already bottlenecked. It won't help adding CUs. Vega 56 performs pretty much the same. You could probably remove more CUs and still not lose much performance.

 

Making a smaller chip would allow AMD to really show their efficiency gains (if they rein in the voltage).

 

So either it's a sub-30 CU that can barely compete with Polaris 10 (which would be mobile Vega repackaged) or perhaps they got a 40ish CU chip that could replace Polaris as their upper mid-range chip. Don't see any other new chip design.

 

It should be noted that Vega Mobile uses 4 GB HBM2 so that's bound to cause problems if that's what they have up their sleeve.

[Jurrunio]

5 hours ago, Trixanity said:

7nm Vega won't see a consumer release.

explain more on that please, I'm interested.

5 hours ago, Trixanity said:

So either it's a sub-30 CU that can barely compete with Polaris 10 (which would be mobile Vega repackaged) or perhaps they got a 40ish CU chip that could replace Polaris as their upper mid-range chip. Don't see any other new chip design.

 

So, no new flag ship designs, that's disappointing. I guess they could stuck a pair of Vega with 40 or 48CU on the same PCB and call that their new flagship, but that doubles the amount of HBM used per card, so it's gonna be even harder to mass produce....

[Jurrunio]

8 hours ago, Valentyn said:

I doubt if AMD is still brave enough to use more than 8GB HBM....

[Trixanity]

7 hours ago, Jurrunio said:

explain more on that please, I'm interested.

So, no new flag ship designs, that's disappointing. I guess they could stuck a pair of Vega with 40 or 48CU on the same PCB and call that their new flagship, but that doubles the amount of HBM used per card, so it's gonna be even harder to mass produce....

Vega on 7nm will be an experiment so to speak. It will be their first 7nm product and will be used to work out the kinks in the new leading process node. Unless the yields are fantastic out of the gate, they'd be wise to recoup their losses through a high price tag.

So it'll be an enterprise/machine learning card. They can sell that for a few thousand dollars.

 

As you can see in the slide you linked (if it's real) they'll also introduce a bunch of new enterprise features. The xGMI thing is supposed to compete with NVLink and Omnipath if I recall correctly and PCI-E 4.0 isn't supposed to arrive in consumer products (instead jumping straight to 5.0 although I think that may still be up in the air).

The only reason I can see them give it a wider release is if the yields are that good and if they feel they really need to release something now instead of waiting for Navi which does seem like it would be the case.

 

Anyway, it's difficult to launch both a new node and a new architecture so using Vega as a test run is a good idea especially when 7nm is really new ground.

 

Yup, it would be pretty silly to release a new big chip that runs into more of the same bottlenecks. It's kneecapped enough as it is. Dual GPU cards are a thing of the past. They're all abandoning it so I can't see them do that due to complexity of the design and the poor support and therefore scaling of performance.

 

I would prefer to get a 40 CU card with 8 gigs of GDDR5X memory but that's not happening. At best it's such a card with HBM2 and probably 4 instead of 8 to fit it into a price range.

 

With that being said:

I'm pretty sure it's the mobile Vega I talked about that they've repackaged into a smaller desktop card. Their previous later releases have been small GPUs like Polaris 11 and 12. And I don't see why other than shortages to not release that on the desktop as well. Expensive memory aside, it's a neat little chip that could replace a bunch of older chips.

[Jurrunio]

8 minutes ago, Trixanity said:

I'm pretty sure it's the mobile Vega I talked about that they've repackaged into a smaller desktop card. Their previous later releases have been small GPUs like Polaris 11 and 12. And I don't see why other than shortages to not release that on the desktop as well. Expensive memory aside, it's a neat little chip that could replace a bunch of older chips.

I'd rather have that in an APU instead. Will most likely push the TDP up to 95w or even more, but it will crush the 1050ti.

[Trixanity]

2 minutes ago, Jurrunio said:

I'd rather have that in an APU instead. Will most likely push the TDP up to 95w or even more, but it will crush the 1050ti.

Might see that on 7nm if it's feasible. 

 

AMD needs to also fill the lower end with decent new cards. They can coexist. I mean they can sell that to everyone including Intel customers and to AMD customers who still need the biggest possible CPU. I also imagine Apple wants to remain with Intel for CPU but wants powerful graphics (relatively speaking) for stuff like their iMac. There's a lot to gain from a chip that size even if it seems somewhat useless when you want it to compete in a higher tier where it's most needed for gaming.

 

In fact I imagine Apple being the target for this chip. It could potentially be put into the MacBook Pro 15 and iMac. That's a lot of money even if Apple tries to squeeze their customers.

[Jurrunio]

1 minute ago, Trixanity said:

Might see that on 7nm if it's feasible. 

 

AMD needs to also fill the lower end with decent new cards. They can coexist. I mean they can sell that to everyone including Intel customers and to AMD customers who still need the biggest possible CPU. I also imagine Apple wants to remain with Intel for CPU but wants powerful graphics (relatively speaking) for stuff like their iMac. There's a lot to gain from a chip that size even if it seems somewhat useless when you want it to compete in a higher tier where it's most needed for gaming.

 

In fact I imagine Apple being the target for this chip. It could potentially be put into the MacBook Pro 15 and iMac. That's a lot of money even if Apple tries to squeeze their customers.

if it's in a card, they have to worry about HBM again. I mean, Vega M equipped intel CPU (forget the name) seems to have sunk in the sea, and that uses HBM.

[Bananasplit_00]

R7 APU perhaps? 

[AresKrieger]

Well this seems to be a mobile or lower power sku, its not the one people are looking forward to which as stated above is Vega 20

 

54 minutes ago, Trixanity said:

Might see that on 7nm if it's feasible. 

Well considering GloFlo 7nm =/= Intel or IBM 7nm it is feasible but ultimately only equivalent to around 9-12nm (don't know the exact number but its in that range most likely 10)

[cj09beira]

2 hours ago, Jurrunio said:

if it's in a card, they have to worry about HBM again. I mean, Vega M equipped intel CPU (forget the name) seems to have sunk in the sea, and that uses HBM.

amd has said before the intel cpu doesn't use vega M

[TheRandomness]

I'd love to see say, a 5120 shader Vega card. Sure, it might put our more heat than a thermonuclear bomb, but it'd be awesome. Or, perhaps, cards with 1280, 1536, 1792, 2048 and 2304 shaders. Or maybe 2560 and 2816 shader cards, to represent the 590 and 590X?

[Jurrunio]

3 minutes ago, cj09beira said:

amd has said before the intel cpu doesn't use vega M

Intel's own Hades Canyon NUC will come first with Kaby Lake G (yes, the Vega M + Intel Kaby Lake HQ chip all-in-one) and is available for pre-order on some sites

https://www.simplynuc.com/hades-canyon/

[asus killer]

weren't vega 11 and 12 supposed to be the replacements for polaris? i remember reading about it last summer.

[cj09beira]

31 minutes ago, TheRandomness said:

I'd love to see say, a 5120 shader Vega card. Sure, it might put our more heat than a thermonuclear bomb, but it'd be awesome. Or, perhaps, cards with 1280, 1536, 1792, 2048 and 2304 shaders. Or maybe 2560 and 2814 shader cards, to represent the 590 and 590X?

we might see things like that with navi, as its supposed to be scalable

[TheRandomness]

8 minutes ago, cj09beira said:

we might see things like that with navi, as its supposed to be scalable

I'd love to imaging 1024 and 2048 shader Navi dies, then just combining them in any way possible Such as two 2048 dies for high end, 2048 and 1024 for mid range, and 2048 for lowish end, and then 1024 and lower for absolute low end. They have so many possibilities with that.

[Trixanity]

3 hours ago, Jurrunio said:

if it's in a card, they have to worry about HBM again. I mean, Vega M equipped intel CPU (forget the name) seems to have sunk in the sea, and that uses HBM.

They'll have to worry about HBM as long as the memory controller only supports HBM. And it seems Vega regardless of form factor (excluding APUs with shared memory controller) will run into that problem. Supposedly there exists lower cost HBM solutions but it comes at the price of capacity and bandwidth. 

2 hours ago, AresKrieger said:

Well this seems to be a mobile or lower power sku, its not the one people are looking forward to which as stated above is Vega 20

 

Well considering GloFlo 7nm =/= Intel or IBM 7nm it is feasible but ultimately only equivalent to around 9-12nm (don't know the exact number but its in that range most likely 10)

GloFo is IBM. IBM paid GloFo to take their foundry business. Yes, you read that correctly. IBM paid GloFo 1.5 billion dollars to get rid of it. Key people from IBM are in charge of the most crucial aspects. 

 

42 minutes ago, TheRandomness said:

I'd love to see say, a 5120 shader Vega card. Sure, it might put our more heat than a thermonuclear bomb, but it'd be awesome. Or, perhaps, cards with 1280, 1536, 1792, 2048 and 2304 shaders. Or maybe 2560 and 2816 shader cards, to represent the 590 and 590X?

Such a big chip would be memory bandwidth starved and still be bottlenecked by the lack of ROPs among other things. GCN can't go any further without a massive rework.

 

43 minutes ago, Jurrunio said:

Intel's own Hades Canyon NUC will come first with Kaby Lake G (yes, the Vega M + Intel Kaby Lake HQ chip all-in-one) and is available for pre-order on some sites

https://www.simplynuc.com/hades-canyon/

I do believe Lisa Su said that the GPU in Kaby Lake G is not in fact Vega Mobile. I mean it's still Vega-based but it's not the same chip. It was designed specifically for Intel.