Vega/Navi Rumors (Updated)

[CatMerc]

Is Polaris 12 a mobile GPU or Polaris on 14LPP (as opposed to LPE)?

Polaris 12 is a smaller Polaris, nothing special. 640 SP's.

All of Polaris is already made on LPP, the rumor Videocardz was talking about was of minimal credibility.

 

[Ajay]

Polaris 12 is a smaller Polaris, nothing special. 640 SP's.

All of Polaris is already made on LPP, the rumor Videocardz was talking about was of minimal credibility.

Someday, I will learn

 

[Mopetar]

Vega11 will be smaller, Vega10 is already 500mm² big. There is no space for a bigger chip, but P10 is 230mm². Vega11 should be in the 300-350mm² region to fill the hole between P10 and V10.

Doesn't that 500 mm^2 include the HBM2 stacks as well though? I thought that actual GPU was smaller. I think they could make a bigger GPU die as they have the same number of shaders as Fury and it seems unlikely that the new features (HBCC, etc.) consume so much space that they could move to having another 16-32 CUs if necessary. I think they just don't have a market for something like that yet, and the process may still be too young to support larger dies, never mind the increased size of the interposer necessary as well.

 

[Krteq]

Doesn't that 500 mm^2 include the HBM2 stacks as well though?

Nope, around 500mm² is for GPU die only.

 

[Valantar]

Nope, around 500mm² is for GPU die only.

I just hope "around" 500 doesn't mean like 520 or something - given that GP102 is 471mm2, that would give Nvidia an advantage in terms of just churning out a significantly bigger die down the road.

 

[lobz]

After the coming Polaris respin, how cheaply do you think AMD can sell the "tuned" up Radeon RX480 for..? Are the yields there, where AMD can go lower on the price for the 480/580..?

$149 bucks..? (-$50).
Taking on the 1080p & 1440p consumers. Leaving Radeon RX Vega brand for VR & 4k.

not gonna happen, unfortunately

 

[Mopetar]

Nope, around 500mm² is for GPU die only.

Is the new stuff like HBCC or whatever else they added taking up that much space or is the density that low, because at 500 mm^2 that puts Vega 10 somewhere around 80-87% of the size of Fiji while having the same number of CUs. We know from Polaris that it's not quite, but close to twice as much possible density on GF 14 nm as you could get at the end of 28 nm, but this is nowhere close.

AMD needs to have massively improved their shader efficiency and utilization or this would be on Bulldozer levels of failure.

 

[Glo.]

After the coming Polaris respin, how cheaply do you think AMD can sell the "tuned" up Radeon RX480 for..? Are the yields there, where AMD can go lower on the price for the 480/580..?

$149 bucks..? (-$50).
Taking on the 1080p & 1440p consumers. Leaving Radeon RX Vega brand for VR & 4k.

Polaris 10 GPUs will cost up to 299$.

 

[JDG1980]

Vega cards are confirmed by AMD to be named "RX Vega".

No, they aren't. They are confirmed to use "Vega" as the retail name. The possible inclusion of the "RX" prefix was inference on the part of various sites (including Anandtech). It might happen, it might not.

 

[Arachnotronic]

No, they aren't. They are confirmed to use "Vega" as the retail name. The possible inclusion of the "RX" prefix was inference on the part of various sites (including Anandtech). It might happen, it might not.

Bro...

Says "Radeon RX Vega" right on the thing.

 

[IllogicalGlory]

Is the new stuff like HBCC or whatever else they added taking up that much space or is the density that low, because at 500 mm^2 that puts Vega 10 somewhere around 80-87% of the size of Fiji while having the same number of CUs. We know from Polaris that it's not quite, but close to twice as much possible density on GF 14 nm as you could get at the end of 28 nm, but this is nowhere close.

AMD needs to have massively improved their shader efficiency and utilization or this would be on Bulldozer levels of failure.

It hits 12.5GFLOPS, so at the very least, clock speeds are way higher. Kinda reminds me of Maxwell vs Kepler. A small increase in shaders/mm^2, but a big boost in clock speed. Pascal vs Maxwell is pretty similar too.

 

[Mopetar]

It hits 12.5GFLOPS, so at the very least, clock speeds are way higher. Kinda reminds me of Maxwell vs Kepler. A small increase in shaders/mm^2, but a big boost in clock speed. Pascal vs Maxwell is pretty similar too.

That doesn't necessarily tell us much beyond a clock speed bump though.If you straight up compared the Fury X to the 980 Ti, the latter only had about 66% of the FLOPs of the former, but generally beat it in most games. If that ratio remained unchanged, it would mean that Vega would only be at the level of the 1080. Of course that's the worst case scenario, but it's still not good.

Maybe they've got a lot of special sauce and we see a huge performance jump and that ratio narrows or AMD even overtakes NVidia in terms of what they can do with their FLOPs. After Ryzen, just about anything seems possible.

 

[IllogicalGlory]

That doesn't necessarily tell us much beyond a clock speed bump though.If you straight up compared the Fury X to the 980 Ti, the latter only had about 66% of the FLOPs of the former, but generally beat it in most games. If that ratio remained unchanged, it would mean that Vega would only be at the level of the 1080. Of course that's the worst case scenario, but it's still not good.

Maybe they've got a lot of special sauce and we see a huge performance jump and that ratio narrows or AMD even overtakes NVidia in terms of what they can do with their FLOPs. After Ryzen, just about anything seems possible.

Just saying, at the very least, it's a fair improvement over the Fury X. I'm personally not all that confident in their arch improvements, but I hope I'm wrong. They claimed to have a lot of arch improvements with Polaris too, none of which added up to much.

We were hoping that the new geometry engine brought them to parity with NV in tessellation, that the new command processor eliminated DX11 overhead and I was hoping that the new multimedia engine would fix idle and video playback power consumption. None of that really happened.

The 980 Ti was actually a bit closer (it had more like 6.8TFLOPS at 1216MHz) if you compared actual clock speeds rather than nominal. Still, the Fury X seems to be stuck fighting it out with the 980 and 1060 in most recent games. Hope Vega isn't massively underutilized like that.

 

[Valantar]

It hits 12.5GFLOPS, so at the very least, clock speeds are way higher. Kinda reminds me of Maxwell vs Kepler. A small increase in shaders/mm^2, but a big boost in clock speed. Pascal vs Maxwell is pretty similar too.

Maxwell and Kepler were on the same process ... If all AMD can get out of the jump from 28nm to 14nm and their biggest architectural update in years is a ~15% shrink in die size and some increased, clocks, they are in trouble.

 

[AtenRa]

My prediction is that small Vega (64 CUs) will be close to 360mm2.

 

[Samwell]

My prediction is that small Vega (64 CUs) will be close to 360mm2.

Give it 44-48 CUs and it might fit, but 64 CUs is big Vega.

Is the new stuff like HBCC or whatever else they added taking up that much space or is the density that low, because at 500 mm^2 that puts Vega 10 somewhere around 80-87% of the size of Fiji while having the same number of CUs. We know from Polaris that it's not quite, but close to twice as much possible density on GF 14 nm as you could get at the end of 28 nm, but this is nowhere close.

AMD needs to have massively improved their shader efficiency and utilization or this would be on Bulldozer levels of failure.

Don't compare it to Fiji. Fiji is a terribly balanced gpu. They took Tonga and have just thrown a ton more shaders in it plus HBM. It's fast in shader workload and bandwidth, but slow in everything else. That's why you see the 480X sometimes fighting with it. P10 is a better comparision with 36 CUs in 238 mm². Double that gpu and you should get 72 CUs in ~450 mm², as bigger gpus have a bit better CU/mm ratios. Now you can add in HBM, which saves some space, but you loose your savings with HBCC and the infinity fabric. Add the tile based approach, which needs bigger caches which takes some space and bigger shaders which have 2xFP16 and 4xInt8 capability and you land at 500mm²

You can have a look at Nvidia too and you'll see the same.
GM200 600mm² vs GP102 470mm². 8 vs 12 Billion transistors.
Same memory interface and rops, so you would expect 50% more shaders as architecture improvements seem to be minor. But Gp102 only has 25% more shaders. Higher clocks, preemption and 4xInt8 capability took the additional space instead of shaders. And GM200 was already a well balanced gpu, which doesn't need to make up these weaknesses. But it still lacks stuff, which Vega will have like 2xFP16 support and other professional features.

 

[Valantar]

Just saying, at the very least, it's a fair improvement over the Fury X. I'm personally not all that confident in their arch improvements, but I hope I'm wrong. They claimed to have a lot of arch improvements with Polaris too, none of which added up to much.

We were hoping that the new geometry engine brought them to parity with NV in tessellation, that the new command processor eliminated DX11 overhead and I was hoping that the new multimedia engine would fix idle and video playback power consumption. None of that really happened.

The 980 Ti was actually a bit closer (it had more like 6.8TFLOPS at 1216MHz) if you compared actual clock speeds rather than nominal. Still, the Fury X seems to be stuck fighting it out with the 980 and 1060 in most recent games. Hope Vega isn't massively underutilized like that.

While I do somewhat feel the same, I think you're a bit overly pessimistic. From what we know, development on Polaris and Vega started out roughly at the same time (or at least they started talking about them at about the same time). That Vega is launching a full year after Polaris then must imply either that
a) It's a far larger departure from GCN than Polaris was (as Polaris was pretty much a die shrink with some added features and overall tweaking of the architecture)
b) It's been pushed back to extract the most performance, or
c) They're really struggling to make it compete

While all three might be true to a degree, I can't quite get myself to believe that c) would be the dominant factor. After all, even with the relatively small tweaks that Polaris brought to the table, they closed the efficiency gap a little, and the feature gap a lot. A year of development beyond that, including process improvements, on an architecture developed in parallel rather than in sequence, should net them some nice improvements unless every engineer they had working on Vega somehow turned into a chimpanzee or something. Of course, this is impossible to tell.

My prediction is that small Vega (64 CUs) will be close to 360mm2.

You know, I'm not entirely opposed to that, although I think 380-400 is more realistic. Fiji was 596mm2 with 4096 SPs/64 CUs. Hawaii was 438mm2 with 2560 (although that includes more FP64 components due to its 1/2 DP compute rate vs. Fiji's 1/16). Polaris is 232mm2 for 2304 SPs (also 1/16 FP64 rate). So for a 10% reduction in SPs from Hawaii, with some added space savings from reduced FP64, they got a 48% die size reduction out of the jump from 28nm to 14nm. CU-to-CU when comparing Hawaii to Polaris, the reduction is still 42%. From Fiji to Polaris it's still a 31% reduction in area per CU. While the reduction from Fiji won't be as drastic due to it already having cut FP64, a direct extrapolation of Polaris 10 to 64 CUs lands at 412mm2 (232/36*64). It seems reasonable to me to expect a slightly smaller die than this, unless the changes to Vega's NCUs are so drastic as to render them incomparable with previous GCN designs. Then again, Hawaii extrapolated to 64 CUs would land us at a 700mm2 die, so the reduction from Hawaii to Fiji was pretty significant too.

Give it 44 CUs and it might fit, but 64 CUs is big Vega.

That would require the Vega NCU to be significantly bigger than a regular GCN CU. Polaris with 44 CUs would barely exceed 280mm2.

(Of course, die sizes can't be extrapolated like I've been doing here, as that doesn't factor in any other components than the SPs themselves. ROPs, TUs and the like also play a part, not to mention memory controllers, encode/decode parts, and so on. Still, it can be useful as a rough approximation).

 

[KompuKare]

That Vega is launching a full year after Polaris then must imply either that
a) It's a far larger departure from GCN than Polaris was (as Polaris was pretty much a die shrink with some added features and overall tweaking of the architecture)
b) It's been pushed back to extract the most performance, or
c) They're really struggling to make it compete

How about
d) Ryzen is way more important and AMD are resource constraint and while Radeon Group engineers are not going to help with polishing Ryzen, the people do process spins, validation, etc. might be busy and not have had much time for Vega.
Or even in terms of launching the product. So if they've had silicon back for months but had to wait until after Ryzen, then with your point a) they've hopefully used that time to polish drivers. If this is the biggest change to GCN then it would make sense that they will need lots of driver work to get the new features working.
Also, while Vega is important if they plan to use it for the Ryzen APUs, then there's plenty of work which they could be going there. The APUs are probably way more important than Ryzen R7, R5 etc. (only server Zen parts would have similar priority).

 

[lobz]

Is the new stuff like HBCC or whatever else they added taking up that much space or is the density that low, because at 500 mm^2 that puts Vega 10 somewhere around 80-87% of the size of Fiji while having the same number of CUs. We know from Polaris that it's not quite, but close to twice as much possible density on GF 14 nm as you could get at the end of 28 nm, but this is nowhere close.

AMD needs to have massively improved their shader efficiency and utilization or this would be on Bulldozer levels of failure.

the stream processors and so the NCUs are much bigger, I'm almost 100% sure

 

[lobz]

Polaris 10 GPUs will cost up to 299$.

that's not gonna happen either, fortunately.

 

[lobz]

My prediction is that small Vega (64 CUs) will be close to 360mm2.

the vega with 64 CUs is big(!!) vega -> the vega 10. small vega will be vega 11, the 'sweet spot' GPU according to AMD. It's starting to be frustrating, why does somebody new come every single day and try to spread this nonsense, that the 64 CU 4096 SP vega 10 is '''''small'''' vega with its close to 500mm2 die size?

 

[lobz]

That would require the Vega NCU to be significantly bigger than a regular GCN CU. Polaris with 44 CUs would barely exceed 280mm2.

they will be much bigger, that's almost 100% certain

edit: formatting

 

[zinfamous]

that's not gonna happen either, fortunately.

Not that I think they can squeeze such performance out of Polaris 10 with this 1 year refresh, but if the 570 or 580 (whatever the new top Polaris becomes) manages 85-95% of 1070, then $299 is more than fair, imo.

Besides, that "up to $299" price target reflects AMD's own words about a year ago, just ahead of Polaris launch. If it happens, it just means AMD were right all along, and this was their goal all along, and further discounts the previous speculation that 480X only launched at that up to $250 price "due to nVidia pricing and performance." Not that any of that speculation isn't true, but none of it isn't not yet true, either.

 

[antihelten]

Bro...

Says "Radeon RX Vega" right on the thing.

To be fair, in that image the "RX" is a suffix to Radeon, not a prefix to Vega as JDG1980 was talking about, so he's not 100% wrong .

You know, I'm not entirely opposed to that, although I think 380-400 is more realistic. Fiji was 596mm2 with 4096 SPs/64 CUs. Hawaii was 438mm2 with 2560 (although that includes more FP64 components due to its 1/2 DP compute rate vs. Fiji's 1/16). Polaris is 232mm2 for 2304 SPs (also 1/16 FP64 rate). So for a 10% reduction in SPs from Hawaii, with some added space savings from reduced FP64, they got a 48% die size reduction out of the jump from 28nm to 14nm. CU-to-CU when comparing Hawaii to Polaris, the reduction is still 42%. From Fiji to Polaris it's still a 31% reduction in area per CU. While the reduction from Fiji won't be as drastic due to it already having cut FP64, a direct extrapolation of Polaris 10 to 64 CUs lands at 412mm2 (232/36*64). It seems reasonable to me to expect a slightly smaller die than this, unless the changes to Vega's NCUs are so drastic as to render them incomparable with previous GCN designs. Then again, Hawaii extrapolated to 64 CUs would land us at a 700mm2 die, so the reduction from Hawaii to Fiji was pretty significant too.

That would require the Vega NCU to be significantly bigger than a regular GCN CU. Polaris with 44 CUs would barely exceed 280mm2.

(Of course, die sizes can't be extrapolated like I've been doing here, as that doesn't factor in any other components than the SPs themselves. ROPs, TUs and the like also play a part, not to mention memory controllers, encode/decode parts, and so on. Still, it can be useful as a rough approximation).

We already have pictures of big Vega (1,2), and based on the size of the HBM2 modules (which is known), the die size can be estimated to be somewhere around 500-550 mm2.

 

[AtenRa]

the vega with 64 CUs is big(!!) vega -> the vega 10. small vega will be vega 11, the 'sweet spot' GPU according to AMD. It's starting to be frustrating, why does somebody new come every single day and try to spread this nonsense, that the 64 CU 4096 SP vega 10 is '''''small'''' vega with its close to 500mm2 die size?

And you know that because what ???