AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

[LordEC911]

Stacked Ram will only improve speed or will we get a lot more capacity ? Seems like we have been at 2-3 gigs for a long time

Both, depending on your definition of a lot more capacity.

 

[eastmen]

Both, depending on your definition of a lot more capacity.

I'd hope to see a doubling of the low end cards hitting 4 to 6 gigs and the mid and high end cards having 6 to 8 gigs at least.

 

[Esrever]

Would they really go with a big gpu first again? Seems like nvidia's approach is better for marketing.

 

[3dilettante]

Both, depending on your definition of a lot more capacity.

If it's first-gen HBM, getting to 4 GB would require four stacks at 128 GB/s bandwidth (some fiddling may come from speed grade) each.

A 2-stack design would get 256 GB/s in bandwidth, which is decent but not difficult to reach with existing memory buses. The capacity is a constrained 2 GB.

A 4-stack would get 512 GB/s, which is 60% above what a consumer Hawaii board can get at 4 GB capacity. The high-end FireStream boards would see a massive drop in capacity with first-gen HBM.

I have not seen mockups going beyond 4 stacks on the interposer. Gen 2 should significantly help with capacity and bandwidth.

 

[RedVi]

4GB would be fine for a high end gaming card for the next year at least. I don't see it as anywhere near as big of a potential bottleneck as 2GB was in 2012.

 

[LordEC911]

If it's first-gen HBM, getting to 4 GB would require four stacks at 128 GB/s bandwidth (some fiddling may come from speed grade) each.

A 2-stack design would get 256 GB/s in bandwidth, which is decent but not difficult to reach with existing memory buses. The capacity is a constrained 2 GB.

A 4-stack would get 512 GB/s, which is 60% above what a consumer Hawaii board can get at 4 GB capacity. The high-end FireStream boards would see a massive drop in capacity with first-gen HBM.

I have not seen mockups going beyond 4 stacks on the interposer. Gen 2 should significantly help with capacity and bandwidth.

Yeah, the density image were throwing me off with 8Hi-stacks.

They have the product page up now...

 

[pjbliverpool]

Stacked Ram will only improve speed or will we get a lot more capacity ? Seems like we have been at 2-3 gigs for a long time

Seems to me that stacked RAM will decrease capacity potential rather than increase it. As 3dilettante says, 4GB would seem to be the first gen limit of HMC for now whereas with the 4Gb GDDR5 chips it's theoretically possible for a card such as the 290x today to come packing 16GB.

The current 2-4GB norm seems to be a commercial decision rather than a technical limitation.

 

[eastmen]

would 2 pools of memory work well . One of stacked ram at very fast speeds and then a slower pool of ram .

 

[psurge]

Why is the density so low? Are they using an older process for the layers in the stack in order to maximize yield? Or is the goal to keep each stack as small as possible, so as to keep required interposer size for a large GPU that wants 4 stacks manageable?

 

[DieH@rd]

I presume they are starting with small chips so that they can guarantee that everything will work fine.

1GB per stack, 4 stacks, 512 GB/s bandwith to the GPU chip, more than good enough for start and for 95% of gamers that are still on single monitor 1080p.

 

[A1xLLcqAgt0qc2RyMz0y]

AMD's next high-end graphics delayed until the second half of 2015

From Pacific Crest Securities:

McConnell and Chung said AMD's next high-end graphics processing unit launch has been delayed until the second half of 2015.

http://news.investors.com/110314-72...etition.htm?ven=yahoocp&src=aurlled&ven=yahoo​

​

 

[AlNom]

Doesn't bode well for 20nm availability.

 

[silent_guy]

Doesn't bode well for 20nm availability.

Was there still anyone left who expected desktop GPUs in 20nm?

I find it hard to believe that AMD will be without anything new for another 8 months. They must have been doing something more than Tonga in the last year...

 

[iMacmatician]

I find it hard to believe that AMD will be without anything new for another 8 months. They must have been doing something more than Tonga in the last year...

I haven't seen any new Iceland rumors lately, but hopefully that GPU will show up before 2015 H2.

 

[AlNom]

Was there still anyone left who expected desktop GPUs in 20nm?

Are they actually skipping straight to 16?

 

[silent_guy]

Are they actually skipping straight to 16?

Sure looks like it. gm204 already shows that Nvidia thought it was worth staying with 28nm, even though 20nm is available for mass production. (Apple)

20nm is a choice between power and speed. Can't have both. And it's only expected to become cheaper per transistor than 28nm by the time 16nm comes around. So it's not even a cost reduction.

 

[AlNom]

Sure looks like it. gm204 already shows that Nvidia thought it was worth staying with 28nm, even though 20nm is available for mass production. (Apple)

20nm is a choice between power and speed. Can't have both. And it's only expected to become cheaper per transistor than 28nm by the time 16nm comes around. So it's not even a cost reduction.

Ah k. Thanks.

Is 16nm not going to be expensive for a while too?

 

[silent_guy]

Is 16nm not going to be expensive for a while too?

I suppose so. But if it's both lower power and faster, at least you get some value for your money.

Like GK110, Pascal was announced at GTC, cost is less of an issue with compute. Wouldn't surprise me one bit if we'd see a same trajectory here: start with compute, have GeForce later, with a bunch units disabled, only introduce the full thing when it's cost effective.

 

[eastmen]

Sure looks like it. gm204 already shows that Nvidia thought it was worth staying with 28nm, even though 20nm is available for mass production. (Apple)

20nm is a choice between power and speed. Can't have both. And it's only expected to become cheaper per transistor than 28nm by the time 16nm comes around. So it's not even a cost reduction.

apple has chips a fraction of the size of what NVidia or amd would need.

20nm will still be worth while since 16nm will be more expensive than 20nm for awhile

 

[3dilettante]

Was there still anyone left who expected desktop GPUs in 20nm?

I find it hard to believe that AMD will be without anything new for another 8 months. They must have been doing something more than Tonga in the last year...

Just putting it out there--but working on something over the last year or so doesn't preclude that the project fails to come to market, not that we have any indication other than the lack of output to go by.

AMD has an unknown number of other projects, such as new semicustom work, continuing work with the consoles, APUs, ARM, wobbling between multiple fab processes, bleeding off engineers and design IP, and a number of initiatives like Mantle and HSA (remember that one?).

In terms of HSA, Carrizo is promised to be the first truly HSA-compliant solution, likely in part due to it having the first GPU from AMD capable of preemption across all work types. If Tonga is the IP level Carrizo uses, then it points to at least the possibility that a number of the poorly-described features AMD barely talked about for the 285 launch are unexposed/incompletely-validated components to what Carrizo needs to have running.

We've already seen a lack of notable design progress on the low-power x86 side, where Puma is at a design level what Jaguar was promised to be when the core was first announced, with power management and turbo capability Jaguar should have had at the start if the necessary characterization and validation had been able to complete.

While I don't know what things AMD has tasked itself with over that time period--such as whether it was working on more things or not, I get the impression that what it has on its plate is greater than or equal to what it can perform.