AMD: Speculation, Rumors, and Discussion (Archive)

[ImSpartacus]

Aaaaww yeah I called it first in this forum (I deserve extra internet points and self-pattings). The two Polaris chips are set to replace Bonaire+Pitcairn and Hawaii.
This is excellent strategy IMO.



His point during the video is that the userbase for the cards that fall within or over the minimum spec for VR is still rather small at 7.5 million users. So for VR to take off and increase the number of possible VR adopters, they need to come up with a cheaper replacement for cards at that performance range, which is where Polaris comes in.
In this case, I think he definitely means lower cost to the consumer (meaning lower than $350), otherwise his whole point wouldn't make sense.

That makes sense. So it looks like Polaris 11 is based on the vr minimum spec and also conveniently replaces the unprofitable Hawaii.

So 2016 (and likely a bit of 2017) will have Polaris 10 holding down the console-tier laptop front, Polaris 11 doing battle at the vr min spec and then Gemini providing a vr "halo" product?

Do you think we'll ever see a dual Polaris 11 part? I know that'll likely conflict with fiji, but vr seems to be a good use case for dual gpu setups and Polaris 11 will likely be fairly efficient (making a dual gpu card easier).

 

[3dilettante]

That makes sense. So it looks like Polaris 11 is based on the vr minimum spec and also conveniently replaces the unprofitable Hawaii.

So 2016 (and likely a bit of 2017) will have Polaris 10 holding down the console-tier laptop front, Polaris 11 doing battle at the vr min spec and then Gemini providing a vr "halo" product?

Do you think we'll ever see a dual Polaris 11 part? I know that'll likely conflict with fiji, but vr seems to be a good use case for dual gpu setups and Polaris 11 will likely be fairly efficient (making a dual gpu card easier).

So a replacement of Hawaii, cheaper than Hawaii, and if AMD is to be believed somewhere around twice the thermal headroom.
My earlier question of what happens when Hawaii is danger-close to much of Fiji's range with all that wiggle room remains.
Gemini looks like it might be the only safe-ish place for Fiji silicon, and only as long as it takes for Polaris 11x2 to be announced.
A dual Polaris 11 would erase Fiji's niche since it could fit dual GPUs under the non-X2 power limits of the current generation.

 

[lanek]

Do you think we'll ever see a dual Polaris 11 part? I know that'll likely conflict with fiji, but vr seems to be a good use case for dual gpu setups and Polaris 11 will likely be fairly efficient (making a dual gpu card easier).

Honestly, dual gpu as Fury X2, 295x2 or their Nvidia counterpart, are not really intended to sold a lot.. So, if they replace it with another one more powerfull ( around the last quarter of 2016 ) the impact will be seriously limited.

 

[Deleted member 13524]

Do you think we'll ever see a dual Polaris 11 part? I know that'll likely conflict with fiji, but vr seems to be a good use case for dual gpu setups and Polaris 11 will likely be fairly efficient (making a dual gpu card easier).

Thanks to LiquidVR, a dual Polaris 11 setup is bound to be quite a bit better at VR than a single Fiji, but I think AMD will avoid releasing a dual-Polaris 11 card so that they don't cannibalize the Fury cards.
However, AMD probably wouldn't stop card manufacturers from releasing their own custom dual-P11 cards with a PLX in the middle, like the Asus' ARES cards.

A dual Polaris 11 would erase Fiji's niche since it could fit dual GPUs under the non-X2 power limits of the current generation.

Well according to Ryan, dual-GPU setups are in a pinch for single-screen gaming because of some games coming out without AFR support and others coming with a 60FPS lock.

 

[ImSpartacus]

Honestly, dual gpu as Fury X2, 295x2 or their Nvidia counterpart, are not really intended to sold a lot.. So, if they replace it with another one more powerfull ( around the last quarter of 2016 ) the impact will be seriously limited.

Yeah, and it was brought to my attention that a dual Polaris 11 card would not only cannibalize fiji cards, but it would also quickly be replaced/superseded by Greenland/Vega10 in 2017ish (?).

Dual gpu cards that aren't based on the halo gpu are always an awkward affair. Maybe vr changes that in the longer term, but it surely won't happen soon enough for a dual Polaris 11 to make sense.

 

[3dilettante]

Thanks to LiquidVR, a dual Polaris 11 setup is bound to be quite a bit better at VR than a single Fiji, but I think AMD will avoid releasing a dual-Polaris 11 card so that they don't cannibalize the Fury cards.

Do you mean cannibalize dual-GPU Fury cards?
What can make a chip that can do just as well as Hawaii in order to serve as a replacement, at a discount, and far less power, not cannibalize the Fury cards that are frequently close to what Hawaii can do anyway?

Well according to Ryan, dual-GPU setups are in a pinch for single-screen gaming because of some games coming out without AFR support and others coming with a 60FPS lock.

Dual-GPU setups were having problems prior to the latest trends in optimization, but it certainly doesn't help.

Dual gpu cards that aren't based on the halo gpu are always an awkward affair. Maybe vr changes that in the longer term, but it surely won't happen soon enough for a dual Polaris 11 to make sense.

Would Fiji be the halo GPU if the architectural and process improvements for a Hawaii replacement come through as advertised?

 

[CarstenS]

In this case, I think he definitely means lower cost to the consumer (meaning lower than $350), otherwise his whole point wouldn't make sense.

Just take into account who he was talking to and his agenda. Of course he'll sell his product in the best possible light to the audience. As does every company representative.

 

[Deleted member 13524]

Just take into account who he was talking to and his agenda. Of course he'll sell his product in the best possible light to the audience. As does every company representative.

And you think selling his product in the best possible light is done by lying about its price point?

 

[CarstenS]

Don't be childish. He did not say anything about the price point of the product, only that IHVs would be able to produce them at lower cost. Which of course is a wording designed to imply a lower price point without saying it explicitly.

 

[Razor1]

And you think selling his product in the best possible light is done by lying about its price point?

They aren't lying, they are being very vague about it, AKA MSRP to be determined. Hell if they can get larger margins why won't they? In recent years looks like AMD fears getting money, but surely that isn't the case

 

[ImSpartacus]

Do you mean cannibalize dual-GPU Fury cards?
What can make a chip that can do just as well as Hawaii in order to serve as a replacement, at a discount, and far less power, not cannibalize the Fury cards that are frequently close to what Hawaii can do anyway?

Nah, a single Fiji-based card. I figure year-old Fiji gets a Grenada-esque refresh for something like ~10% more performance than 2015's Fiji.

Meanwhile, Polaris 11 aims to make the minimum VR spec as economical as possible, so I assume that means cut-down (i.e. cheaper) Polaris 11 performs roughly like a 290 (min Oculus Rift spec), maybe a smidge better.

So we've got a Fiji with maybe +10% performance compared to a dual GPU card that's packing two GPUs in the ballpark of the 290-290X. So that's effectively a hot-clocked Fury X versus a 295X2. If TechPowerUp's famously simple illustrations are to be trusted, then see below and add like 10% to the Fury X's score:

I call that close enough to the 295X2 to cause cannibalization concerns, so no dual Polaris 11 cards.

This kind of thing is unavoidable when you try to make a dual CPU card out of a mid-tier GPU. That's why we only see dual GPU cards from the halo GPUs of a particular generation.

 

[kalelovil]

Edit: delete

 

[lanek]

Nah, a single Fiji-based card. I figure year-old Fiji gets a Grenada-esque refresh for something like ~10% more performance than 2015's Fiji.

Meanwhile, Polaris 11 aims to make the minimum VR spec as economical as possible, so I assume that means cut-down (i.e. cheaper) Polaris 11 performs roughly like a 290 (min Oculus Rift spec), maybe a smidge better.

So we've got a Fiji with maybe +10% performance compared to a dual GPU card that's packing two GPUs in the ballpark of the 290-290X. So that's effectively a hot-clocked Fury X versus a 295X2. If TechPowerUp's famously simple illustrations are to be trusted, then see below and add like 10% to the Fury X's score:

I call that close enough to the 295X2 to cause cannibalization concerns, so no dual Polaris 11 cards.

This kind of thing is unavoidable when you try to make a dual CPU card out of a mid-tier GPU. That's why we only see dual GPU cards from the halo GPUs of a particular generation.

Please, dontt use TechpowerUP as plotting point, witth all the respect i have for Wizard , their reviews, tests, are seriously going wroong since some years. If you want to have a point, use your own test here .

 

[gamervivek]

The zauba samples were going around for the same price as Hawaii's, so it's not a big surprise.

There have been rumors that there are three chips, the biggest one isn't called Greenland anymore(quite credible) but is named Vega10(less credible) and that it is not the same architecture as the other two chips(least credible). Thus Koduri could go around saying that they have two Polaris chips for 2016 without telling the full truth that the third chip is due for 2016 as well. nvidia will most probably launch their big(>=400mm2 or thereabouts) chip this year itself and it will be around 40-50% faster than the 980Ti. So if AMD wait for 2017 for their big chip then nvidia will end up stealing their lunch and dinner and everything in between.

As for a souped up Fiji for tiding over AMD till the new chips are around, it doesn't sound likely. Fiji is at 1050Mhz and is the biggest chip. It'd require engineering something akin to the 4870->4890 to make it go any higher. Driver improvements could change things a bit, it's apparent that GCN3 got the short end of the stick compared to GCN2 and GCN1 for drivers and their mantle driver, which reflects in poor DX12 performance of Fiji compared to Hawaii, is a good example. But even then it will likely end up 10-15% faster than a 980Ti and still be beaten by the ubiquitous OCed 980Ti models.

 

[Kaotik]

There have been rumors that there are three chips, the biggest one isn't called Greenland anymore(quite credible) but is named Vega10(less credible) and that it is not the same architecture as the other two chips(least credible).

Greenland is supposedly part of MCM-chip, if one leaked slide is real. Others have suggested it would be part of "monster APU" which could have been misinterpreted thinking "one chip" would mean APU, while it's acually MCM

 

[Alessio1989]

Any news about rasterization capabilities? A working ROVs support would be really, really appreciated, even more the CR.

 

[ImSpartacus]

As for a souped up Fiji for tiding over AMD till the new chips are around, it doesn't sound likely. Fiji is at 1050Mhz and is the biggest chip. It'd require engineering something akin to the 4870->4890 to make it go any higher. Driver improvements could change things a bit, it's apparent that GCN3 got the short end of the stick compared to GCN2 and GCN1 for drivers and their mantle driver, which reflects in poor DX12 performance of Fiji compared to Hawaii, is a good example. But even then it will likely end up 10-15% faster than a 980Ti and still be beaten by the ubiquitous OCed 980Ti models.

So you're thinking that it wouldn't be possible to meaningfully refresh Fiji even if AMD let out some slack with respect to power/noise/heat?

I mean, Hawaii feels like it was probably pushed harder than it should've been pushed to begin with (recall the 290's late fan speed adjustment to eek out additional performance). And yet, Grenada came onto the scene and was a fair improvement beyond Hawaii (the 390 effectively replaced the 290X and the 390X edged towards the 980 and the 380X was able to be released now that the 290 was retired). AMD kept pushing that poor chip and it kept performing (and heating homes in the winter).

But in the event that Fiji isn't able to be pushed harder, do you think AMD will retire it entirely? I mean, AMD has been forced to reuse basically all of its GPUs for at least one additional year since the first round of GCN stuff.

 

[gamervivek]

Fiji could probably get 8GB HBM2 version, but has been pretty much tapped out on core clocks. So barely an improvement.

290X wasn't that far off a 980 if the reference coolers weren't used. Grenada came with launch driver that made great tessellation improvement which pushed it even closer and its core clock is same as Fury X. Pretty unlikely that the bigger chip would get a similar clock boost.

Reusing chips happened with 28nm and it was a straightforward progression of increasing die sizes. Doesn't make much sense now. A 300-350mm2 chip from nvidia would beat it rather comfortably, so AMD would have to be pretty suicidal to consider keeping Fury at the top for whole of 2016.

 

[gamervivek]

Greenland is supposedly part of MCM-chip, if one leaked slide is real. Others have suggested it would be part of "monster APU" which could have been misinterpreted thinking "one chip" would mean APU, while it's acually MCM

A recent slide? Have been seeing too many posts about using multiple dies on an interposer recently at anandtech and other forums..

 

[Razor1]

Fiji could probably get 8GB HBM2 version, but has been pretty much tapped out on core clocks. So barely an improvement.

290X wasn't that far off a 980 if the reference coolers weren't used. Grenada came with launch driver that made great tessellation improvement which pushed it even closer and its core clock is same as Fury X. Pretty unlikely that the bigger chip would get a similar clock boost.

Reusing chips happened with 28nm and it was a straightforward progression of increasing die sizes. Doesn't make much sense now. A 300-350mm2 chip from nvidia would beat it rather comfortably, so AMD would have to be pretty suicidal to consider keeping Fury at the top for whole of 2016.

Depends on AMD's strategy to return to profitability, which is much more important than to have a halo product for one or two quarters........ Midrange is much more important since that is the honey pot of volume sales along with margins. We can try to infer from slides all day long but what goes on behinds the scenes shows more to the health of AMD than temporary and short term applause.

PS the APU slide, that seems to fit along with what AMD has been saying about their financial strategy they have taken thus far for CPU and GPU lines along with what Raja have stated so far.