Not that it would be in any way unexpected but that roadmap seems to indicate that they'll skip 10nm. They can offer enough perf per $ gains on 14/16nm with Vega due to lower costs and HBM2, instead of going 10nm very early. By mid 2018, 7nm should be available so hard to imagine Navi would go 10nm.
I am not betting, i am concluding. As for 7nm, it could be ready well before that, mid 2018 wouldn't be that early. From your tone I do understand that you are a TSMC hater (apparently those exist for some reason ) but the way TSMC does it, the 7nm transition could be quick.TSMC is pairing processes and reusing the vast majority of tools. 20 with 16 and 10 with 7nm. Additionally , 10nm brings only 15% speed gains over 16 and that's not worth it in GPU when you can just go wider. So yeah, i am concluding that they will skip 10nm and Navi will be on 7nm,whenever it arrives.
I doubt they are a "TSMC Hater", it is simply that process transitions have been steadily slowing down, year by year. Whether you realize it or not, 5nm is the lower limit of electron-tunneling based designs. At that point, it is not possible to prevent an electron from existing on both sides of the channel, purely based on its size.
I don't think whatever TSMC brands as "7 nm" is going to have anything to do with the 5x10^-9 meter limit you're referring to. Probably they will take whatever is ready to go in that time frame and then push it out. Fabs are business, they need a regular product refresh, and they'll likely have one.
Below the 7nm channel width (around 5nm) it becomes impossible to ignore the effects of quantum tunneling effects. We will either need a radical new transistor design, or a bunch of quantum physicists designing our standard cell libraries.
Either way, process scaling is slowing down, and there is no way that any process from TSMC (or any other fab) will make it to high enough yields to be useful by the end of 2018 for Navi.
I agree, AMD historically in the past has trailed right on the heels of intel's cpus, and for a good discount\performance ratio. It's a good buisness model, and they do it well. I know exactly what you are gonna say, that they are bleeding money out the A$$, but I like that they are continuing to try and be competitive even at lower price points.
So yes, I don't see them dumping a ton of R&D into skipping 10mm and going straight to 7mm, it just doesn't seem like it makes sense financially, and honestly, I think it makes MORE sense for them to take little steps, rather than a jump forward. It's too big of a gamble for them to do that.
Well I mean if you look at what they have done with FX.... They left FX on 32nm and are now going to 14nm. Because its cheaper that way to squeeze as much as possible out of each node. Which is why it would make sense to skip 10 and wait for 7.
TSMC current roadmap assumed 7nm "risk production" at the end of 2017. My guess is Apple puts the A10 on 16nmFF+, 2017's A11 on 10nm, and then MAYBE (this would be really aggressive) 2018's A12 on 7nm. But A12 production has to start about 9 months before iPhones ship to build up inventory, so that's crazy tight. Which means more likely 2019's A13 gets 7nm.
Reason Apple is relevant in all this is that Apple so far gets first dibs at everything new and good at TSMC. (It would not surprise me if this is, by now, contractual --- Apple pays part of developing the new process, but then gets 6 months exclusive use or whatever.) I think it would be astonishing if AMD (or anyone else of significance) shipped a chip, especially a chip the size of serious GPUs, before Apple gets its shipments out.
Unless something dramatic changes in the industry, GPU will never be using the leading node anytime soon. Those capacity are reserved for Mobile SoC only. So even in 2018 it is likely to be 10nm.
TSMC 7nm at the end of 2018 will very likely be Apple SoC only.
This is a common misconception with process size. Just because the process has a process size of 7nm, this doesn't mean that every part of the design is or even can be the minimum feature size. Going from one FinFET process to the next means that an old 14nm might not even need to be shrunk in any way to be produced under a 7nm process. Now, of course they will optimize what they can, but AMD is a shining example of designers continually improving designs on the same process node.
Ah, that "new to tech" optimism. You clearly haven't been following manufacturing process rollouts long enough. Also, 2016 (14 nm GPU launch date) + 2 years (typical time between nodes) = 2018.
Unless 7nm is something incredibly watered down and thus abnormally quick to market... you couldn't be more wrong.
If we're to see high power 7nm necessary for GPUs in 2018 that means low power products would ship next year, and that's likely not the case.
As for 'TSMC haters' they're mostly just the people who remember that the firm had trouble making deadlines well back into planar shrinks and think that they're full of it whenever they make claims of early launches of FinFET technology at problematic process sizes.
Wow, AMD must be raking in a lot if cash recently and spending it all on R&D to be able to come up with so many new architectures in such a short period of time. Very nice.
It's just an incremental design process. Polaris is a 16nm design sticking with the GDDR5 interface. Vega will likely be the high end replacements for the Fury line with the new 16nm architecture from Polaris and an HBM2 memory controller. Expect any alterations to Polaris from Vega to trickle down to the mid-year cards in 2017 (possibly HBM1/2). With Navi replacing the Vega Fury line in late 2017/ early 2018 with either an improved HBM2 interface, or the replacement to HBM2.
I'm reading elsewhere that Polaris was designed to support GDDR5/HBM1/HBM2.
"Earlier this year AMD confirmed to us that while it is still committed to the High Bandwidth Memory technology it co-invented with SK Hynix and brought to the market last year, Polaris was designed to be compatible with both HBM/HBM2 and GDDR5 memory standards. Technical marketing lead at AMD Robert Hallock explained that the company has the flexibility to use either technology as needs arise. This means that either memory technology can be employed where it makes sense."
Press conference said GDDR5 for Polaris. Vega is most likely Polaris with HBM and some efficiency tweaks (think Tonga).
I'm going off of previous release cycles where AMD puts out the affordable consumer models first (260X and 290X), and then launches enthusiast models later (295X2). June is way too soon to replace Fury. This is why Vega will be later this year, early next.
It does look like the Polaris generation is split into Vega and Polaris. I wonder if it's a possibility that AMD will go without a high-end GPU till Vega? Clearly for a high-end Finfet GPU 4GB is not enough any more, and GDDR5X is too slow. I'm guessing it was their plan to deliver the Polaris high-end GPU with HBM2 but due to HBM2 being delayed now that has been moved to Vega. Pretty sure they did similar reschedules with Fiji due to 20nm being bust and HBM delayed.
Nvidia is probably doing the same. GP104 with G5X this year, GP100 with HBM2 next to take on Vega.
GDDR5X is at least somwhat competitive to HBM1, ~448GB/s vs 512GB/s, of course HBM2 doubles that again.
If they remain limited to 4GB on HBM1, one can certainly hope AMD pushes for a (at least close to high-end) with GDDR5X, as 4GB just doesn't cut it for a 2016 GPU.
Bandwidth wasn't the primary reason HBM was conceived. It's power budget. 5X at it's lowest voltage uses 15% more power than HBM. It also can only scale up 2 voltage tiers. Obviously those will make it look worst in terms of power budget vs HBM.
Looks absolutely sane. They very badly need competitive mobile (and low/mid end) GPU's. If they get this release right these will serve as that for much of the duration of 14/16nm.
Worth remembering that claimed power efficiency gain - 2.5 times. Even if its 'only' ~2x, in practice they should be able to cover the high end quite well at ~150w until really big 14/16nm stuff rolls up.
Don't worry, these are just codenames like the star names of Phenom/Phenom II era. End customers will be buying 4xx, 5xx etc. series cards and not Polaris, Vega etc., just like they bought Phenom, Phenom II or Athlon II rather than Toliman, Deneb, Regor CPUs ...
Wait so no HBM2 , no GDDR5X but just plain GDDR5 , well that is a suprize. Since it is a perf/watt slide that gain for vega is then mainly the memory?
Personally i was kinda hoping to see cards with 8gb(>=6GB) becoming standard in 2016. So wondering : -they trying to get it out really early to the market ? -something went wrong? -Hope for AMD that NVIDIA doesn't suprize them and getting HBM2/GDDR5X out on their sub 20nm.
My guess would be that the market shifted under their feet and they're trying to respond. When work on HBM began, VR wasn't a thing yet. With the advent of VR and a general gaming focus on improving textures and supersampling, the 4gb limit that HBM offered suddenly seemed restrictive. From there, AMD also decided it didn't want to invest in competing technologies, so it stepped back to good old reliable DDR5 to bring cards with lots of vRAM to the market.
I'd still be really surprised if Polaris didn't feature ANY SKUs with HBM; HBM is great for situations where you are space or bandwidth constricted.
I too hope this doesn't mean Nvidia gets the drop on them with HBM2; AMD really can't afford to lose that way to Nvidia right now.
These are going to be quite high volume cards, and at least eventually moderately priced as well. HBM doesn't suit that much. They might throw it in for a top end card or something, but you can imagine not.
Very likely that NV will get HMB2 out first from this roadmap - they need that for the big pascal based Tesla's and they've got contracts for those shipping relatively soon. How soon that becomes relevant for the consumer market maybe another matter!
"Polaris 10 is AMD's "console-class" graphics card for thin-and-light notebooks. The company expects that graphics chips and cards using Polaris 10 silicon will be able to deliver as much as two times the performance per watt of Nvidia's GeForce GTX 950. The company says that it demonstrated Polaris 11 in December of last year. If that's the case, a Polaris 11 chip running Star Wars Battlefront at 1080p and 60 FPS drew about 90W at the wall."
If Polaris 11 is the smaller GPU, then the 'console class' Polaris 10 would end up with a TDP higher than 90W, which doesn't make sense for notebooks.
That would make a lot of sense. Smaller dies (better yield) with very fast interconnects through an interposer could logically work in the same way as a monolithic big GPU. With exploding mask costs at smaller nodes reducing the number of different dies would bring major cost benefits.
Insofar as there is "new design" coming, maybe it is targeted at AI? IF AI is going to be the breakout market over the next ten years, it would make sense to tweak to GPU to better accommodate whatever AI wants.
Or to put it differently, if AMD thinks they can survive purely on doing VR well, and leaving nV to pick up the AI market (with all those data warehouses), they're, IMHO, making a nasty mistake.
Most likely this. The Gddr5 is the cheapest available "highend" memory (low end ddr3...) and Gddr5+ will come too late. Also the HBM2 is coming too late, so that is why Vega goes to 2017. Interesting to see how much common Polaris and Vega have architectual. Are they the "same" with different memory architecture, or is Polaris tweaked 3xx series with smaller node and Vega tweaked Fiji with smaller node. It would make sense, because the different memory can affect the chip design in larger scale. Well end of the year or first half of 2017 will show.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
54 Comments
Back to Article
jjj - Tuesday, March 15, 2016 - link
Not that it would be in any way unexpected but that roadmap seems to indicate that they'll skip 10nm.They can offer enough perf per $ gains on 14/16nm with Vega due to lower costs and HBM2, instead of going 10nm very early.
By mid 2018, 7nm should be available so hard to imagine Navi would go 10nm.
emn13 - Tuesday, March 15, 2016 - link
So you're betting 7nm will be ready for mass-release in mid 2018? That sounds optimistic to me...jjj - Tuesday, March 15, 2016 - link
I am not betting, i am concluding. As for 7nm, it could be ready well before that, mid 2018 wouldn't be that early.From your tone I do understand that you are a TSMC hater (apparently those exist for some reason ) but the way TSMC does it, the 7nm transition could be quick.TSMC is pairing processes and reusing the vast majority of tools. 20 with 16 and 10 with 7nm.
Additionally , 10nm brings only 15% speed gains over 16 and that's not worth it in GPU when you can just go wider.
So yeah, i am concluding that they will skip 10nm and Navi will be on 7nm,whenever it arrives.
SaberKOG91 - Tuesday, March 15, 2016 - link
I doubt they are a "TSMC Hater", it is simply that process transitions have been steadily slowing down, year by year. Whether you realize it or not, 5nm is the lower limit of electron-tunneling based designs. At that point, it is not possible to prevent an electron from existing on both sides of the channel, purely based on its size.saratoga4 - Tuesday, March 15, 2016 - link
I don't think whatever TSMC brands as "7 nm" is going to have anything to do with the 5x10^-9 meter limit you're referring to. Probably they will take whatever is ready to go in that time frame and then push it out. Fabs are business, they need a regular product refresh, and they'll likely have one.SaberKOG91 - Tuesday, March 15, 2016 - link
Below the 7nm channel width (around 5nm) it becomes impossible to ignore the effects of quantum tunneling effects. We will either need a radical new transistor design, or a bunch of quantum physicists designing our standard cell libraries.Either way, process scaling is slowing down, and there is no way that any process from TSMC (or any other fab) will make it to high enough yields to be useful by the end of 2018 for Navi.
ImSpartacus - Tuesday, March 15, 2016 - link
It sounds like 10nm would be another 20nm mess. Where are you finding this info?Shadow7037932 - Tuesday, March 15, 2016 - link
I seriously doubt we would have mass production at 7nm by 2018. We've only got 14nm for mainstream mass production recently.Kvaern2 - Tuesday, March 15, 2016 - link
When was the last time TSMC delivered a process on schedule?Michael Bay - Tuesday, March 15, 2016 - link
In 90s!jjj is insane anyway.
Spunjji - Wednesday, March 16, 2016 - link
Explain to me the difference between an optimistic conclusion based on faulty inferences and a bet? xDsmilingcrow - Monday, March 21, 2016 - link
$$$$$$$$$$$$$$$$$$iAMD - Tuesday, March 15, 2016 - link
i'm already working on 0.00000000000001 nmdsraa - Tuesday, March 15, 2016 - link
I agree, AMD historically in the past has trailed right on the heels of intel's cpus, and for a good discount\performance ratio. It's a good buisness model, and they do it well. I know exactly what you are gonna say, that they are bleeding money out the A$$, but I like that they are continuing to try and be competitive even at lower price points.So yes, I don't see them dumping a ton of R&D into skipping 10mm and going straight to 7mm, it just doesn't seem like it makes sense financially, and honestly, I think it makes MORE sense for them to take little steps, rather than a jump forward. It's too big of a gamble for them to do that.
Jimster480 - Tuesday, March 15, 2016 - link
Well I mean if you look at what they have done with FX.... They left FX on 32nm and are now going to 14nm. Because its cheaper that way to squeeze as much as possible out of each node. Which is why it would make sense to skip 10 and wait for 7.name99 - Wednesday, March 16, 2016 - link
TSMC current roadmap assumed 7nm "risk production" at the end of 2017.My guess is Apple puts the A10 on 16nmFF+, 2017's A11 on 10nm, and then MAYBE (this would be really aggressive) 2018's A12 on 7nm. But A12 production has to start about 9 months before iPhones ship to build up inventory, so that's crazy tight. Which means more likely 2019's A13 gets 7nm.
Reason Apple is relevant in all this is that Apple so far gets first dibs at everything new and good at TSMC. (It would not surprise me if this is, by now, contractual --- Apple pays part of developing the new process, but then gets 6 months exclusive use or whatever.)
I think it would be astonishing if AMD (or anyone else of significance) shipped a chip, especially a chip the size of serious GPUs, before Apple gets its shipments out.
iwod - Tuesday, March 15, 2016 - link
Unless something dramatic changes in the industry, GPU will never be using the leading node anytime soon. Those capacity are reserved for Mobile SoC only. So even in 2018 it is likely to be 10nm.TSMC 7nm at the end of 2018 will very likely be Apple SoC only.
jayfang - Tuesday, March 15, 2016 - link
SoC's have GPU's too, thus 7nm GPU designsNot just a smart comment => if Navi is scalable then it may also target Mobile SoC's
SaberKOG91 - Tuesday, March 15, 2016 - link
This is a common misconception with process size. Just because the process has a process size of 7nm, this doesn't mean that every part of the design is or even can be the minimum feature size. Going from one FinFET process to the next means that an old 14nm might not even need to be shrunk in any way to be produced under a 7nm process. Now, of course they will optimize what they can, but AMD is a shining example of designers continually improving designs on the same process node.III-V - Tuesday, March 15, 2016 - link
Ah, that "new to tech" optimism. You clearly haven't been following manufacturing process rollouts long enough. Also, 2016 (14 nm GPU launch date) + 2 years (typical time between nodes) = 2018.Unless 7nm is something incredibly watered down and thus abnormally quick to market... you couldn't be more wrong.
OrenjiJuusu - Tuesday, June 7, 2016 - link
If we're to see high power 7nm necessary for GPUs in 2018 that means low power products would ship next year, and that's likely not the case.As for 'TSMC haters' they're mostly just the people who remember that the firm had trouble making deadlines well back into planar shrinks and think that they're full of it whenever they make claims of early launches of FinFET technology at problematic process sizes.
D. Lister - Tuesday, March 15, 2016 - link
Wow, AMD must be raking in a lot if cash recently and spending it all on R&D to be able to come up with so many new architectures in such a short period of time. Very nice.varad - Tuesday, March 15, 2016 - link
It hardly costs any money to make slides ;). Snide remarks aside, here's hoping we have healthy competition in the GPU world.D. Lister - Tuesday, March 15, 2016 - link
:) of course.toxicfiend1957 - Tuesday, March 15, 2016 - link
thumbs up i agreeZingam - Tuesday, March 15, 2016 - link
Slide-making is the top paid job ever.SaberKOG91 - Tuesday, March 15, 2016 - link
It's just an incremental design process. Polaris is a 16nm design sticking with the GDDR5 interface. Vega will likely be the high end replacements for the Fury line with the new 16nm architecture from Polaris and an HBM2 memory controller. Expect any alterations to Polaris from Vega to trickle down to the mid-year cards in 2017 (possibly HBM1/2). With Navi replacing the Vega Fury line in late 2017/ early 2018 with either an improved HBM2 interface, or the replacement to HBM2.DanNeely - Tuesday, March 15, 2016 - link
I'm reading elsewhere that Polaris was designed to support GDDR5/HBM1/HBM2."Earlier this year AMD confirmed to us that while it is still committed to the High Bandwidth Memory technology it co-invented with SK Hynix and brought to the market last year, Polaris was designed to be compatible with both HBM/HBM2 and GDDR5 memory standards. Technical marketing lead at AMD Robert Hallock explained that the company has the flexibility to use either technology as needs arise. This means that either memory technology can be employed where it makes sense."
http://wccftech.com/amd-unveils-polaris-vega-navi-...
With HBM2 delayed until Vega, I'd assume we'll see some higher end Polaris parts with HBM1. Hopefully with 8gb instead of only 4.
SaberKOG91 - Tuesday, March 15, 2016 - link
Press conference said GDDR5 for Polaris. Vega is most likely Polaris with HBM and some efficiency tweaks (think Tonga).I'm going off of previous release cycles where AMD puts out the affordable consumer models first (260X and 290X), and then launches enthusiast models later (295X2). June is way too soon to replace Fury. This is why Vega will be later this year, early next.
trane - Tuesday, March 15, 2016 - link
It does look like the Polaris generation is split into Vega and Polaris. I wonder if it's a possibility that AMD will go without a high-end GPU till Vega? Clearly for a high-end Finfet GPU 4GB is not enough any more, and GDDR5X is too slow. I'm guessing it was their plan to deliver the Polaris high-end GPU with HBM2 but due to HBM2 being delayed now that has been moved to Vega. Pretty sure they did similar reschedules with Fiji due to 20nm being bust and HBM delayed.Nvidia is probably doing the same. GP104 with G5X this year, GP100 with HBM2 next to take on Vega.
nevcairiel - Tuesday, March 15, 2016 - link
GDDR5X is at least somwhat competitive to HBM1, ~448GB/s vs 512GB/s, of course HBM2 doubles that again.If they remain limited to 4GB on HBM1, one can certainly hope AMD pushes for a (at least close to high-end) with GDDR5X, as 4GB just doesn't cut it for a 2016 GPU.
Despoiler - Tuesday, March 15, 2016 - link
Bandwidth wasn't the primary reason HBM was conceived. It's power budget. 5X at it's lowest voltage uses 15% more power than HBM. It also can only scale up 2 voltage tiers. Obviously those will make it look worst in terms of power budget vs HBM.ImSpartacus - Tuesday, March 15, 2016 - link
That's likely. We've already heard rumors of Greenland being renamed Vega 10. Now that we know the "10" gpu is the beefiest, it makes sense.Qwertilot - Tuesday, March 15, 2016 - link
Looks absolutely sane. They very badly need competitive mobile (and low/mid end) GPU's. If they get this release right these will serve as that for much of the duration of 14/16nm.Worth remembering that claimed power efficiency gain - 2.5 times. Even if its 'only' ~2x, in practice they should be able to cover the high end quite well at ~150w until really big 14/16nm stuff rolls up.
prtskg - Wednesday, March 16, 2016 - link
Wasn't the claimed efficiency 2X but it was w.r.t. nano (As we saw in testing). W.r.t. other GCN gpus efficiency improvement is 2.5X.r3loaded - Tuesday, March 15, 2016 - link
Ugh c'mon, not another product naming scheme! I was just getting comfortable with R9 2XX and R9 3XX.Arnulf - Tuesday, March 15, 2016 - link
Don't worry, these are just codenames like the star names of Phenom/Phenom II era. End customers will be buying 4xx, 5xx etc. series cards and not Polaris, Vega etc., just like they bought Phenom, Phenom II or Athlon II rather than Toliman, Deneb, Regor CPUs ...SaberKOG91 - Tuesday, March 15, 2016 - link
Agreed. Polaris 10 seems to be performance optimized, while Polaris 11 is power optimized. Very much like the Fury X and Fury Nano, respectively.Arnulf - Tuesday, March 15, 2016 - link
Wouldn't really surprise me if AMD picked this naming scheme to mess with Nvidia ... e.g. Polaris vs. Pascal, Vega vs. Volta, Navi vs. Newton ...Ammaross - Tuesday, March 15, 2016 - link
+1 point for you sir!plopke - Tuesday, March 15, 2016 - link
Wait so no HBM2 , no GDDR5X but just plain GDDR5 , well that is a suprize.Since it is a perf/watt slide that gain for vega is then mainly the memory?
Personally i was kinda hoping to see cards with 8gb(>=6GB) becoming standard in 2016. So wondering :
-they trying to get it out really early to the market ?
-something went wrong?
-Hope for AMD that NVIDIA doesn't suprize them and getting HBM2/GDDR5X out on their sub 20nm.
Eschaton - Tuesday, March 15, 2016 - link
My guess would be that the market shifted under their feet and they're trying to respond. When work on HBM began, VR wasn't a thing yet. With the advent of VR and a general gaming focus on improving textures and supersampling, the 4gb limit that HBM offered suddenly seemed restrictive. From there, AMD also decided it didn't want to invest in competing technologies, so it stepped back to good old reliable DDR5 to bring cards with lots of vRAM to the market.I'd still be really surprised if Polaris didn't feature ANY SKUs with HBM; HBM is great for situations where you are space or bandwidth constricted.
I too hope this doesn't mean Nvidia gets the drop on them with HBM2; AMD really can't afford to lose that way to Nvidia right now.
Qwertilot - Tuesday, March 15, 2016 - link
No need to overcomplicate things :)These are going to be quite high volume cards, and at least eventually moderately priced as well. HBM doesn't suit that much. They might throw it in for a top end card or something, but you can imagine not.
Very likely that NV will get HMB2 out first from this roadmap - they need that for the big pascal based Tesla's and they've got contracts for those shipping relatively soon. How soon that becomes relevant for the consumer market maybe another matter!
prtskg - Wednesday, March 16, 2016 - link
Since both Samsung and Hynix show q3 for volume production of HBM2, I don't think any 'reasonably priced' gpu will feature it till q4.extide - Tuesday, March 15, 2016 - link
Where are you hearing no GDDR5X? I bet we will see GDDR5X on Polaris 10, at least!samlebon2306 - Tuesday, March 15, 2016 - link
"Where are you hearing no GDDR5X?"Because GDDR5X will get produced in mass only this summer, so AMD has no control here.
edzieba - Tuesday, March 15, 2016 - link
"I have also been told that Polaris 11 is the smaller of the Polaris GPUs, so at this point it’s reasonable to assume the same for Vega."That's interesting. Over at Techreport (http://techreport.com/blog/29851/radeon-pro-duo-sp...
"Polaris 10 is AMD's "console-class" graphics card for thin-and-light notebooks. The company expects that graphics chips and cards using Polaris 10 silicon will be able to deliver as much as two times the performance per watt of Nvidia's GeForce GTX 950. The company says that it demonstrated Polaris 11 in December of last year. If that's the case, a Polaris 11 chip running Star Wars Battlefront at 1080p and 60 FPS drew about 90W at the wall."
If Polaris 11 is the smaller GPU, then the 'console class' Polaris 10 would end up with a TDP higher than 90W, which doesn't make sense for notebooks.
narlzac85 - Tuesday, March 15, 2016 - link
Could Navi possibly be a new multi-GPU design of Vega with >1 GPU sharing the HBM interposer and memory (shared memory design)?'scalability' and 'next-gen memory'...seems to fit.
MrSpadge - Tuesday, March 15, 2016 - link
That would make a lot of sense. Smaller dies (better yield) with very fast interconnects through an interposer could logically work in the same way as a monolithic big GPU. With exploding mask costs at smaller nodes reducing the number of different dies would bring major cost benefits.name99 - Wednesday, March 16, 2016 - link
Insofar as there is "new design" coming, maybe it is targeted at AI?IF AI is going to be the breakout market over the next ten years, it would make sense to tweak to GPU to better accommodate whatever AI wants.
Or to put it differently, if AMD thinks they can survive purely on doing VR well, and leaving nV to pick up the AI market (with all those data warehouses), they're, IMHO, making a nasty mistake.
RobATiOyP - Tuesday, March 15, 2016 - link
HBM1 was limitted to 4GiB (which would have seemed like a good first target a few years ago), that's one reason HBM2 is required already.asimov1979 - Tuesday, March 15, 2016 - link
I think Vega will be the true succesor of Fiji, while Polaris will have the rest of the stack.haukionkannel - Tuesday, March 15, 2016 - link
Most likely this. The Gddr5 is the cheapest available "highend" memory (low end ddr3...) and Gddr5+ will come too late. Also the HBM2 is coming too late, so that is why Vega goes to 2017.Interesting to see how much common Polaris and Vega have architectual. Are they the "same" with different memory architecture, or is Polaris tweaked 3xx series with smaller node and Vega tweaked Fiji with smaller node.
It would make sense, because the different memory can affect the chip design in larger scale.
Well end of the year or first half of 2017 will show.
Mat3 - Tuesday, March 15, 2016 - link
These marketing guys are way too optimistic about process shrinks in the near future.