[Ars] AMD confirms high-end Polaris GPU will be released in 2016

[PPB]

You don't need those specs to know there's a huge gap in their lineup.

A small die of Polaris 10 needs a chip twice that size, before getting a chip of ~300mm2 mid-range caliber.

There's no way they will be stupid enough to do a dual Polaris 10 SKU, because CF is still iffy (you think NV's GimpWorks is going to make it easier for AMD to support CF moving forward?!).

Polaris 11 has to be twice harvested to fall into the $300 segment.

Since they stated only 2 Polaris chips, they are missing at the real high-end. Their only option would be Dual Polaris 11 SKU. Which is sub-optimal due to reliance on CF support. If I were to go CF again, I would only do so with 2 huge dies, not 2x mid-range.

I guess its to be expected given the new node woes. AMD trying to produce a massive die on the new node is suicidal.

You can harvest polaris 11 to 33% less shaders and have your x2 polaris 10 sku.

Still to this day, we have a 60% perf delta between 960 and 970. Nv never cared for a 960ti altought they had the 980m salvaged die to use for filling the gap.

I see amd is going after the market trends for once, i think that is a good thing.

 

[parvadomus]

Here is what I am guessing

Polaris 11 - 4096 sp, GCN 4th gen, 128 ROPs, 8GB HBM2, 2048 bit HBM2 bus, 512 GB/s, 180-200w.

Will have 3 salvage SKUs. 3584, 3072 and 2048 sp. The 2048 sp would be a heavily salvaged (50% disabled) GPU and would be sold for USD 350. The improved shader efficiency and architectural improvements should bring it close to GTX 980/R9 390X at roughly 100w.

I think AMD will go GDDR5X for both chips, I didnt know that GDDR5 memory version wasnt that fast.

So for me: Polaris 11 - 4096 sp, GCN 4th gen, 128 ROPs, 8GB GDDR5X, 512 bit GDDR5X bus, 512 GB/s, 150-200w.

It would be stupid to waste on HBM2 tech for such a "small" die, and knowing nvidia will probably go with the same memory tech for that die size. Giving Nvidia´s bigger customer base, I dont think AMD can afford competing with a higher cost product vs Nvidia @ same segment.
Polaris 10 will probably be a 256-bit part.

 

[USER8000]

Polaris 10 is probably a 128 bit part. The GTX960 does one enough at 1080P using a 128 bit memory controller.

 

[Glo.]

Lineup of AMD GPUs:

Fury Nano X2 - 8192 GCN cores HBM1.
Polaris 11 - 6144 GCN cores HBM2. Less than 200W TDP.
Fury Nano - 4096 GCN cores HBM1
Polaris 10 - 2048 GCN cores 384 bit, GDDR5. Under 75W TDP.

 

[Techhog]

If the specs provided by Lorien are correct, then we could see single Polaris 10 up to $250, dual Polaris 10 up to $499 and Polaris 11 from $650 with Dual Polaris 11 up to $1500.

I may buy a R9 390 at $300 today after all.

If this is supposed to mean multi-die, then you can throw all of this out of the window right now.

 

[wege12]

Lineup of AMD GPUs:

Fury Nano X2 - 8192 GCN cores HBM1.
Polaris 11 - 6144 GCN cores HBM2. Less than 200W TDP.
Fury Nano - 4096 GCN cores HBM1
Polaris 10 - 2048 GCN cores 384 bit, GDDR5. Under 75W TDP.

If your suggested specs for Polaris 11 are correct, this chip is going to be something else! (assuming at least 128 ROPS are included this gen)

Makes my Fury X feel outdated already......

 

[tential]

Hard to imagine they would be selling the power hungry Hawaii/Tonga in the 14nm era... insanity.

Polaris low end
Fiji mid range
Polaris high end

How about that?

 

[Silverforce11]

We won't see Fiji again I don't think. It's huge, with interposer and HBM tech that would only delay production of HBM2 SKUs.. recall, the bottleneck was at UMC, the guys who did the TSV & Interposer stacking!

Samsung just announced mass production of HBM2. Hynix is gearing up. So someone's gotta be ordering lots of HBM2 chips, it can only be AMD & NV (which may be delayed).

Polaris 11 I strongly believe is a HBM2 only chip. AMD can get two of these onto a card with comfy TDP for HPC markets, along with Samsung's 4GB HBM2 module, 4 of those on an interposer, 16GB vram, makes for a great HPC part.

The problem with that approach @PPB, is the 33% harvested Polaris 11 with HBM2 is unlikely to fit in the mid-range sweet spot on price (AMD specifically raised cost as an issue with using HBM2), the 970/390 killer segment.

If they don't show up to market with a compelling GPU around $300 to $350, they are going to get screwed. The 970 is the most popular GPU according to Steam's latest survey. There's more 970s than all of AMD's other recent GPUs!!

 

[maddie]

Polaris low end
Fiji mid range
Polaris high end

How about that?

If true, you can get around the following.
Polaris low end 35-50W 2 & 4 GB
Fiji mid-range 175W 4 GB
Polaris high end 150W 8 & 16 GB [assuming 300mm^2]

Your mid-range is using more power than your high end.
With a customer base highly sensitized to power use over the past 2 years, how will the mid-range Fiji model compete against an equivalent performing Nvidia part using 100W? I can see the criticism already.

Think about this:
The purpose of the halo cards although profitable in their own right, is to stimulate other sales. Why have a halo if you have nothing else except the low end. Also most of the development cost would have been all the IP blocks redesigning, learning about 14nm, etc. The additional cost to develop a mid-range die would have been a small additional cost compared to what had been spent already.
Why would AMD forgo spending that additional mid-range development sum and risk loosing most of the middle ranges sales?

 

[Glo.]

We won't see Fiji again I don't think. It's huge, with interposer and HBM tech that would only delay production of HBM2 SKUs.. recall, the bottleneck was at UMC, the guys who did the TSV & Interposer stacking!

Samsung just announced mass production of HBM2. Hynix is gearing up. So someone's gotta be ordering lots of HBM2 chips, it can only be AMD & NV (which may be delayed).

Polaris 11 I strongly believe is a HBM2 only chip. AMD can get two of these onto a card with comfy TDP for HPC markets, along with Samsung's 4GB HBM2 module, 4 of those on an interposer, 16GB vram, makes for a great HPC part.

The problem with that approach @PPB, is the 33% harvested Polaris 11 with HBM2 is unlikely to fit in the mid-range sweet spot on price (AMD specifically raised cost as an issue with using HBM2), the 970/390 killer segment.

If they don't show up to market with a compelling GPU around $300 to $350, they are going to get screwed. The 970 is the most popular GPU according to Steam's latest survey. There's more 970s than all of AMD's other recent GPUs!!

Fiji on 28 nm with HBM will cost less to produce than Polaris on 14 nm with HBM2. Much, MUCH less.

 

[railven]

Fury Nano has already been brought down to the $500 price point. It being a binned full Fiji makes you think they either losing money hand over fist on it, or it is generating some revenue at that price point.

I also wouldn't be surprised if they have another excess chip situation and we see Fiji get the Hawaii->Grenada treatment.

Is HBM2 compatible with HBM1? Ie the memory controllers, the interposer, the overall design? Slap some harvested HBM2 chips onto a Fiji interposer, bam.

I think we'll see some kind of Frakenstein Fiji. Either they get it running on GDDR5(X) or possibly go with HBM1/HBM2 and just say screw margins.

AMD is going to get it's face kicked in on the top end. THe market already shows they are willing to wait and pay extra for Nvidia. And the bottom market should be well covered by Polaris 11.

The middle has been AMD's saving grace and they can't afford to lose it.

 

[Silverforce11]

Fiji on 28 nm with HBM will cost less to produce than Polaris on 14 nm with HBM2. Much, MUCH less.

You think a Fury Nano would be very profitable at $300?

How about when faced against NV's GP104 harvested SKU, the 970 successor, with ~100W vs Nano's 175W? Using much cheaper GDDR5 I bet too.

Suicidal to pit a 28nm product against next-gen uarch on 16nm FF. Exactly like they've been doing on the CPU side, out-dated uarch, out-dated node vs Intel.. didn't go well.

 

[ShintaiDK]

Agree, with 14/16 you make new products or you go home empty handed.

 

[Silverforce11]

The only way this is going to work out for them is Polaris 11 offers exceptional performance and perf/w, allowing them to jack up the price and still be able to sell it.

Basically there won't be a mid-range SKU at ~$300. It'll move to $400, for harvested Polaris 11. Top end Polaris 11 (still mid-range chip) is going to be north of $650...

 

[NTMBK]

Perhaps for mid-range they will make a Fury/Fury Nano refresh with HBM2? Increasing volumes should make HBM and interposers cheaper, so maybe they could get the cost down.

 

[railven]

You think a Fury Nano would be very profitable at $300?

How about when faced against NV's GP104 harvested SKU, the 970 successor, with ~100W vs Nano's 175W? Using much cheaper GDDR5 I bet too.

Suicidal to pit a 28nm product against next-gen uarch on 16nm FF. Exactly like they've been doing on the CPU side, out-dated uarch, out-dated node vs Intel.. didn't go well.

Didn't stop them this gen. And for all accounts those were bang for buck and MUST buys.

 

[Glo.]

You think a Fury Nano would be very profitable at $300?

How about when faced against NV's GP104 harvested SKU, the 970 successor, with ~100W vs Nano's 175W? Using much cheaper GDDR5 I bet too.

Suicidal to pit a 28nm product against next-gen uarch on 16nm FF. Exactly like they've been doing on the CPU side, out-dated uarch, out-dated node vs Intel.. didn't go well.

at 500$ Nano is profitable, and taking into account that 100$ goes to... sellers.

TSV's are currently 10$ cheaper than they were at the start of production. Everything goes down with pricing.

And there is one more thing. Fiji device ID is already in OS X directly linked to Mac Pro Framebuffer. The problem is: It will not be high end part in that computer. So they have where to sell the GPUs.

 

[maddie]

You think a Fury Nano would be very profitable at $300?

How about when faced against NV's GP104 harvested SKU, the 970 successor, with ~100W vs Nano's 175W? Using much cheaper GDDR5 I bet too.

Suicidal to pit a 28nm product against next-gen uarch on 16nm FF. Exactly like they've been doing on the CPU side, out-dated uarch, out-dated node vs Intel.. didn't go well.

Agree, with 14/16 you make new products or you go home empty handed.

Agree fully.

Most of the posts in this thread reminds me of times when scientific paradigms change. The old guard go through convoluted explanations using traditional concepts to explain events.

Koduri says reclaim the top end, not compete at the top end. Fighting words indeed.
They are stressing power use as a big factor for them going forward. This should eliminated reuse of 28nm.

 

[Glo.]

I genuinely hope, that I am wrong and Fury Nano will not be used as part on AMD lineup.

I really do.

 

[MeldarthX]

I genuinely hope, that I am wrong and Fury Nano will not be used as part on AMD lineup.

I really do.

It shouldn't be as they've annouced 1.3dp and hdmi 2.0a support. Fury doesn't have either.....

 

[MrTeal]

Fury Nano has already been brought down to the $500 price point. It being a binned full Fiji makes you think they either losing money hand over fist on it, or it is generating some revenue at that price point.

I also wouldn't be surprised if they have another excess chip situation and we see Fiji get the Hawaii->Grenada treatment.

Is HBM2 compatible with HBM1? Ie the memory controllers, the interposer, the overall design? Slap some harvested HBM2 chips onto a Fiji interposer, bam.

I think we'll see some kind of Frakenstein Fiji. Either they get it running on GDDR5(X) or possibly go with HBM1/HBM2 and just say screw margins.

AMD is going to get it's face kicked in on the top end. THe market already shows they are willing to wait and pay extra for Nvidia. And the bottom market should be well covered by Polaris 11.

The middle has been AMD's saving grace and they can't afford to lose it.

Really, they could just rebrand Hawaii again and cut the cost of it to fill in that gap, or just leave it as is and slash price. Set the MSRP on the 390 to $229 and the 390X to $279 and call it a day. You could even drop it back to 4GB of RAM and clock it back to 900MHz to get power consumption down and lower board costs. Getting GDDR5 onto Fiji would require a pretty substantial redesign of the die; you'd think those engineering resources would be much better spent getting a new midrange chip taped out and ready for when capacity and yields make sense.

 

[Arachnotronic]

Just show this on the net,

http://videocardz.com/58169/nvidia-to-launch-geforce-gtx-980mx-and-gtx-970mx



Ehm any commends on that ?? are those 28nm or 16nm ??

If AMD really does transition to 14/16nm in power-sensitive areas like notebooks, then this could be a pretty big win for AMD.

 

[tential]

If true, you can get around the following.
Polaris low end 35-50W 2 & 4 GB
Fiji mid-range 175W 4 GB
Polaris high end 150W 8 & 16 GB [assuming 300mm^2]

Your mid-range is using more power than your high end.
With a customer base highly sensitized to power use over the past 2 years, how will the mid-range Fiji model compete against an equivalent performing Nvidia part using 100W? I can see the criticism already.

Think about this:
The purpose of the halo cards although profitable in their own right, is to stimulate other sales. Why have a halo if you have nothing else except the low end. Also most of the development cost would have been all the IP blocks redesigning, learning about 14nm, etc. The additional cost to develop a mid-range die would have been a small additional cost compared to what had been spent already.
Why would AMD forgo spending that additional mid-range development sum and risk loosing most of the middle ranges sales?

I think a top down release is feasible. I think it's odd many people think 2 chips automatically = 4 products....

 

[Techhog]

You don't need those specs to know there's a huge gap in their lineup.

A small die of Polaris 10 needs a chip twice that size, before getting a chip of ~300mm2 mid-range caliber.

There's no way they will be stupid enough to do a dual Polaris 10 SKU, because CF is still iffy (you think NV's GimpWorks is going to make it easier for AMD to support CF moving forward?!).

Polaris 11 has to be twice harvested to fall into the $300 segment.

Since they stated only 2 Polaris chips, they are missing at the real high-end. Their only option would be Dual Polaris 11 SKU. Which is sub-optimal due to reliance on CF support. If I were to go CF again, I would only do so with 2 huge dies, not 2x mid-range.

I guess its to be expected given the new node woes. AMD trying to produce a massive die on the new node is suicidal.

See, all of this high-end talk only matters if you're naive enough to believe that we'll see consumer GP100 this year. In reality, we likely won't even see a Titan.

 

[Techhog]

I think a top down release is feasible. I think it's odd many people think 2 chips automatically = 4 products....

I really can't see more than 5 cards with two chips if one is really small. If we were talking about a 200mm2 chip and a 300mm2 chip, then I could see 6 cards (enough for a full lineup without having to make a weak dual-GPU.