AMD Polaris architecture – GCN 4.0

[AllBodies]

it would be cheaper to design a smaller chip then artificially limit yields.

I read an industry theory that AMD plan to go for smaller dies in future, and then stick multiple dies on an interposer for their high-end chips. This due to larger chips having exponentially worse yields (i.e. 2x size chip has much less than half the yield).

It'll be interesting if they do this. And we'll get a strong indicator of this if the next consoles have 2+ gpu's.

 

[D.P.]

It seems to me that unless the 1080 is a lot more than 25% more powerful compared to a TX, then that statement is unlikely to be true.

Such increases of performance happened before and that was without a process shrink. What is there to be surprised about here? Only thing surprising to me is nvidia not charging more for the 1070

I think a large performance difference between big Polaris and the 1080 is quite possible just given the difference in chip sizes, transistors and design goals. They look to be aimed at different markets. So AMD could well be worried without an answer to the 1080.

HOWEVER, its sounds strange to me that AMD are surprised that Nvidia are releasing a mid-sized chip at this time. Did AMD really think Nvidia were 6 months late, or that Nviida would be releasing a 750ti type card first? Doesn't sound likely, they must have known for a long time. I mean or starters the rumored AMD GPUs jump from a small chip in Polaris to a big chip in vega, where the heck is the mid-sized 350-400mm^2 chip? How were they expecting to cover the performance gap?


i really have no idea. I would love for the current Poalris rumours to be wrong, or for there to be some unknown bigger Polaris chip coming, maybe that is what will come in October and now AMD want to push it sooner? Or maybe AMD had to cut the GPU R&D budget and there will will be a big gap between Vegas and Polaris, I surely hope not

 

[AllBodies]

I think a large performance difference between big Polaris and the 1080 is quite possible just given the difference in chip sizes, transistors and design goals. They look to be aimed at different markets. So AMD could well be worried without an answer to the 1080.

HOWEVER, its sounds strange to me that AMD are surprised that Nvidia are releasing a mid-sized chip at this time. Did AMD really think Nvidia were 6 months late, or that Nviida would be releasing a 750ti type card first? Doesn't sound likely, they must have known for a long time. I mean or starters the rumored AMD GPUs jump from a small chip in Polaris to a big chip in vega, where the heck is the mid-sized 350-400mm^2 chip? How were they expecting to cover the performance gap?


i really have no idea. I would love for the current Poalris rumours to be wrong, or for there to be some unknown bigger Polaris chip coming, maybe that is what will come in October and now AMD want to push it sooner? Or maybe AMD had to cut the GPU R&D budget and there will will be a big gap between Vegas and Polaris, I surely hope not

Possibly the plan was to not compete with the 1080, but match/beat the 1070? And now they're worried they can't do that because the 1070 is also faster than the TX?

Again just speculating, I certainly hope the P10 can match the 1070, or get within 5% for £250, or something.

 

[Perfect_Chaos]

If Vega is really coming in October then i see that is likely a much better choice coming from a 980 Ti, if it's true i would wait for that than look at the 1080 i think.

 

[Mtom]

I think a large performance difference between big Polaris and the 1080 is quite possible just given the difference in chip sizes, transistors and design goals. They look to be aimed at different markets. So AMD could well be worried without an answer to the 1080.

HOWEVER, its sounds strange to me that AMD are surprised that Nvidia are releasing a mid-sized chip at this time. Did AMD really think Nvidia were 6 months late, or that Nviida would be releasing a 750ti type card first? Doesn't sound likely, they must have known for a long time. I mean or starters the rumored AMD GPUs jump from a small chip in Polaris to a big chip in vega, where the heck is the mid-sized 350-400mm^2 chip? How were they expecting to cover the performance gap?


i really have no idea. I would love for the current Poalris rumours to be wrong, or for there to be some unknown bigger Polaris chip coming, maybe that is what will come in October and now AMD want to push it sooner? Or maybe AMD had to cut the GPU R&D budget and there will will be a big gap between Vegas and Polaris, I surely hope not

The surprised part in the article is plain stupid i think. They know far better what the other side do than we are.
Possible reason could be that HBM producton is going better than expected.

 

[Mauller]

The original slides showed vega hovering around the years end. So it is no surprise really. They could have been a bit more cautious with their release due to the supply of parts when they released the second slide showing Vega in 2017. But they could have now fixed that issue and decided to release with the original plan.

Considering Global foundaries is supposed to be doing far better than expected with 14nm LPP yields.

Apparently Fury's are turning up in lower numbers, so they could be diverting fiji production for S9300's and pro duo's, while they replace fiji with Vega on the desktop.

 

[CAT-THE-FIFTH]

AMD can't just magically push forward the launch within in a few days,unless the cards are ready for sale this year already??

 

[D.P.]

I read an industry theory that AMD plan to go for smaller dies in future, and then stick multiple dies on an interposer for their high-end chips. This due to larger chips having exponentially worse yields (i.e. 2x size chip has much less than half the yield).

It'll be interesting if they do this. And we'll get a strong indicator of this if the next consoles have 2+ gpu's.

Its interesting, and it makes sense from a HW perspective but there has to be a monumental shift in the software. There are also likely more intelligent ways of doing this by having separate modules of there GPU that can be combined without duplicating all the same pieces, e.g. HVEC codes etc.

But that doesn't really help AMD in the short term if they have nothing between Polaris and vega.

 

[Mtom]

AMD can't just magically push forward the launch within in a few days,unless the cards are ready for sale this year already??

Well they most likely were designed together with Polaris, so we can assume the chips are well on the works, but Hynix HBM will be available in mass in Q3, so likely the memory is the limiting part.

 

[D.P.]

If Vega is really coming in October then i see that is likely a much better choice coming from a 980 Ti, if it's true i would wait for that than look at the 1080 i think.

Definitely but October is a long way off and that is a vague rumour. People buying a 1080 know full well bigger chips are to come by the end of the year.

 

[Mauller]

AMD can't just magically push forward the launch within in a few days,unless the cards are ready for sale this year already??

If it is a HBM2 shortage issue, they can easily have all of the other parts ready to go and awaiting assembly.

 

[Perfect_Chaos]

Either way is a good chance Vega is coming late this year?

 

[Perfect_Chaos]

Definitely but October is a long way off and that is a vague rumour. People buying a 1080 know full well bigger chips are to come by the end of the year.

I wasn't aware either company would bring out the highest end cards near the end of the year, i thought most were saying next year.

 

[D.P.]

AMD can't just magically push forward the launch within in a few days,unless the cards are ready for sale this year already??

No, but they might have been wanting to delay until nearer the end of the year and instead will try to push things faster.

Take GP100, working samples are back form TSMC and are working well but they dont plan to release until end of the year when yields and HBM2 are better available. But they could trickle out cards a few months sooner if HBM2 were available. I'm sure Vega is not that far behind.

But I don't put much weight in this rumor.

 

[D.P.]

Either way is a good chance Vega is coming late this year?

I wasn't aware either company would bring out the highest end cards near the end of the year, i thought most were saying next year.

Neither have been precise, end of this year or early next year.

 

[AlamoX]

maybe AMD just got some unexpected cash, like 300mil here and 300mil there, and they decided they can kick it up a notch ?

 

[FoxEye]

Either way is a good chance Vega is coming late this year?

Don't know because AMD are doing their whole "wall of silence" thing.

Because if there's a sure-fire way to get people excited about your product, it's to... say nothing (?) and let your competition release first and still... say nothing (?) So all we have are rumours and speculation, and mostly those aren't positive, making us quite pessimistic about AMD's new cards. AMD watch all this and try to keep our hopes alive by... saying nothing. So now we all mostly are just waiting to buy up those 1070s whilst AMD keep silent and absolutely on pain of death don't do anything that could be remotely considered good PR.

AMD logic!

 

[AllBodies]

Its interesting, and it makes sense from a HW perspective but there has to be a monumental shift in the software. There are also likely more intelligent ways of doing this by having separate modules of there GPU that can be combined without duplicating all the same pieces, e.g. HVEC codes etc.

But that doesn't really help AMD in the short term if they have nothing between Polaris and vega.

Yeah very true.

On the software front, that's why I mentioned the consoles. If we're ever going to get complete multi-gpu support, it'll have to take the consoles having 2+ gpu's.

 

[D.P.]

maybe AMD just got some unexpected cash, like 300mil here and 300mil there, and they decided they can kick it up a notch ?

I don't get your logic. A cash injection doesn't make you surprised at the competition releasing a mid-end part, nor does it let you suddenly shift release schedules on a whim. The biggest stumbling blocks for Vegas and GP100 is HBM2 availability and 16nm yields, money doesn't help drastically here. the process will mature, engineering will improve yields and HBM2 is ready when it ready.

 

[pmc25]

Well they most likely were designed together with Polaris, so we can assume the chips are well on the works, but Hynix HBM will be available in mass in Q3, so likely the memory is the limiting part.

It seems Vega has a host of new hardware features too, so firmware and drivers will be an additional consideration. It's not supposed to be as big a jump as Fiji / Tonga to Polaris, but it's meant to be much more different in architectural and feature terms than Hawaii / Grenada to Fiji / Tonga. It'll be a significant jump. According to AMD, Polaris' memory controller supports HBM2 already .. Vega isn't just Polaris with an HBM2 memory controller, it's considerably more.

I expect their plan is to hard launch it at CES '17. Whether they stick to that largely depends on HBM2 availability and price, I suspect.

I'd still be really surprised if the biggest Vega is bigger than 450-475mm2. My guess is something in the 425-440mm2 range. Their architectural lead over Pascal ought to be enormous at that point, and they're using a cheaper process with far higher potential production .. I expect them to take full advantage of that with much smaller dies than NVIDIA even in the super high end, and reap the rewards in much lower costs and greater yields.