[WccfTech] 16 nm node delayed

[Mondozei]

So the inevitable has happened.

And yes, as was obvious from the getgo, 20 nm GPUs are going to happen, for AMD at least. NV is the open question now. GM200 on 28 nm could be enough to give AMD a run for it. (And if it is, that's really sad and alarming on so many levels, because that would merely confirm the Intelification of the GPU space).

But I hope not. Either way, 16 nm is delayed and all the early space is taken up by the mobile SoCs.

One thing the story doesn't say though: 14 nm. I'm guessing AMD is holding its guns on that one, but we could(probably should) see early 2016 for 14 nm GPUs for AMD because they have preferential access to Global Foundries in a way that NV doesn't have to TSMC.

AMD will have a de facto node leadership vs Nvidia for the next 16 months or so. Let's hope that's enough to bring it into technical parity with NV under Lisa Su. She's intellectually brilliant and an engineer at heart. She was also deep-knee in the AMD/Samsung&GloFo deal for 14 nm. So she has already proven her skills. Will it be enough? Let's hope so, because if the GPU space becomes like the desktop CPU space, we can kiss progress goodbye.

 

[Black Octagon]

Shame that WFFCTech can't reveal their sources. The site has such a pathetic record. Most of their accurate articles are the ones that source someone else from the public domain. Their insider intelligence has been shown to be rubbish on more than just a few occasions

 

[S.H.O.D.A.N.]

Uhm, so far, AMD has been rather sucessful with providing performance comparable to NVidia's on the same node. I'm not sure where the idea that NVidia can beat AMD on an older node comes from, unless you expect the 20nm to be a very poor performer.

 

[Enigmoid]

AMD 300 series on 20nm seems to match with the previous leaks on the 390/380 performance and power consumption. Did anyone truely believe TSMC and the other foundries were on point with their time schedule? IMO take everyone's time frame and add at least 25% to it. That seems to be the case recently for everyone.

 

[RussianSensation]

The article hints at a fully unlocked 285X and a refresh of 290X (295X?) as stop-gaps between Caribbean Islands in May of 2015. What's strange about that is AMD could have released a fully unlocked Tonga and a 7-10% faster 290X a long time ago. In addition, AMD would have known a long time ago that 20nm node is wafer constrained. OTOH, if they will release 285X and 295X for only 5 months, it seems awfully risky and dangerous to have do that which makes me think R9 300 series, if made on 20nm, is delayed past May 2015 if the refresh rumours are true. Also, we haven't heard any concrete details about R9 285/290X refreshes, so it doesn't look like they'll have those ready for January unless AMD is now as secretive as Apple. I am prepared to be surprised.

I really hope that GM200 and 390/390X deliver the 50% increase over 780Ti that many have been waiting for by summer of 2015 at the latest.

 

[NTMBK]

Thanks for putting WCCF in the title so I didn't need to read the story. That site's track record on this is utter garbage.

Without reading, I bet they ripped off the S|A story from last week (while screwing up major details) and didn't cite sources.

 

[tviceman]

I really hope that GM200 and 390/390X deliver the 50% increase over 780Ti that many have been waiting for by summer of 2015 at the latest.

Lower your expectations to guarantee being surprised. Just kidding.

I foresee very little performance difference between GM200 and whatever AMD's 20nm flagship GPU is.... Power consumption numbers and die sizes will be interesting (and a focal point of many arguments, I'm sure). I'd love to see the financials behind the yields and wafer costs of a ~550mm2 28nm chip vs. a ~350-75mm2 20nm chip, but I doubt that will ever happen. More than anything now, I'm as interested in the tech itself as I am the performance. I guess I'm getting old.

 

[raghu78]

AMD has publicly stated that 28nm will be the node on which the majority of their products are built in 2015. I am quite sure that AMD's next gen GPUs are manufactured at GF 28SHP with 2.5D stacking and HBM. I have given my reasons in the posts below with more detail at semiaccurate

http://forums.anandtech.com/showthread.php?t=2409989&page=10
http://semiaccurate.com/forums/showpost.php?p=226455&postcount=30

With AMD confirming Carrizo and Carrizo-L at 28nm with a single package we do not need any more confirmation that 20nm is a no go for high end GPUs. Anybody who believes that 20nm is suitable for high performance GPUs with 250W TDP when AMD has decided to skip it for a < 100 sq mm (if built at 20nm) low power SoC suitable for notebooks and tablets (Carrizo-L) is seriously clueless. Other than a low power and low volume mobile ARM A57 SoC I do not expect any other chip from AMD at 20nm. AMD will make the jump from 28nm to 14nm FINFET for APUs and GPUs in 2016.

http://www.anandtech.com/show/8742/amd-announces-carrizo-and-carrizol-next-gen-apus-for-h1-2015

Interestingly the situation is so bad that even just 6 months back AMD talked about 20nm pin compatible Puma+ and ARM A57 SoCs and here we are in late 2014 and they have completely backtracked on their announcement. Carrizo-L is 28nm and the ARM A57 SoC does not find any mention on AMD's 2015 mobility product roadmap.

http://www.anandtech.com/show/7989/...bridge-pincompatible-arm-and-x86-socs-in-2015

Other than Apple, Qualcomm and Samsung and few FPGA vendors I do not expect any other fabless chip vendor to use 20nm. Samsung has started FINFET production and will slowly ramp volumes in H1 2015 with a steep curve in H2 2015. Samsung expects 30% of 12 inch wafer production at 14nm FINFET by end 2015.

http://aod.teletogether.com/sec/20141030/2014Q3_script.pdf (page 12)

TSMC will ramp 16FF+ in Q3 2015 and the ramp will be steep with Q4 2015 FINFET wafer revenue at 8- 9%.

http://www.tsmc.com/uploadfile/ir/quarterly/2014/3Vs3J/E/TSMC 3Q14 transcript.pdf (page 17 and 18)

Given how Apple and Qualcomm are hogging leading edge capacity and that Samsung uses their leading edge capacity for the Exynos chips, other companies like AMD or Nvidia cannot expect any meaningful allocation before Q4 2015 or Q1 2016. Nvidia's 16FF+ GPUs (manufactured at TSMC) could make it to market earliest by late Q1 2016. AMD's 14nm FINFET GPUs with 2.5D stacking and HBM2 manufactured at GF should be here by late Q2 2016.

wccftech is a horrible click bait site who speculate on other speculative articles. they have zero credibility and do not perform any diligence. they are literally a bottom of the barrel internet trash site.

 

[RussianSensation]

Lower your expectations to guarantee being surprised. Just kidding.

I foresee very little performance difference between GM200 and whatever AMD's 20nm flagship GPU is.... Power consumption numbers and die sizes will be interesting (and a focal point of many arguments, I'm sure). I'd love to see the financials behind the yields and wafer costs of a ~550mm2 28nm chip vs. a ~350-75mm2 20nm chip, but I doubt that will ever happen. More than anything now, I'm as interested in the tech itself as I am the performance. I guess I'm getting old.

As raghu stated, I don't think 20nm is a shoe in for 300 series. It could be the 2nd case that AMD went huge die on 28nm and ran into yields, temperatures or power consumption issues when trying to get clock speeds good enough to hit their performance targets. Either way we slice it, NV is going to bring out 960 next month and AMD will be 4 months behind. By end of March 2015, AMD's 300 series will be 6 months behind similar to 5850/5870 vs. 470/480, at which point every extra month delay would be very costly if they are stuck on the same 280/280X/285/290/290X series. 960 has the potential to wipe out all of their non-290 sales. Are they going to drop 290 to $199 to combat 960/970? We know that people will pay a premium for NV with lower performance so even if a $250 960 is 5-10% slower than a 290, it's game over for the 290. Realistically since 200 series is now old tech, uses more power, it can't be priced similar to NV even with the same performance.

Surely AMD has ran into some issues, maybe even HBM constrained, because when 960 enters, things will go from bad to ugly. Their mobile dGPU market share will get wiped out by GM206 entering laptops.

 

[Phynaz]

Wccf needs to hire Theo Valach to complete the circle.

 

[Arachnotronic]

Anybody who believes that 20nm is suitable for high performance GPUs with 250W TDP when AMD has decided to skip it for a < 100 sq mm (if built at 20nm) low power SoC suitable for notebooks and tablets (Carrizo-L) is seriously clueless.

Don't tell that to the FPGA vendors. I think Xilinx was boasting that it had built the world's highest transistor count chip on the 20nm process. And those FPGAs consume serious power.

 

[Mondozei]

Thanks for putting WCCF in the title so I didn't need to read the story. That site's track record on this is utter garbage.

Without reading, I bet they ripped off the S|A story from last week (while screwing up major details) and didn't cite sources.

Not really. People make generic attacks but have no facts backing up them.

Wccftech's record is pretty similar to most of the industry's. Often they have been right, especially in the GPU space. If you disagree, do provide facts for those disagreements, I got plenty of material to back up my assertions. (Note: Not saying they are always correct, but their record is not worse than the industry average, often leading the rest to a story).

SA is hilarious. They are taking money to do the same job as Wccftech, and often do a worse job. Feel bad for the suckers who pay to read them.

AMD has publicly stated that 28nm will be the node on which the majority of their products are built in 2015.

Sure, and this is not negated by anything. Most of AMD's products are non-dGPUs. Lisa Su has stated multiple times that they are pushing 20 nm in 2015 for selected GPUs(yes she mentioned graphics specifically), and that only makes sense for their cash cows, i.e. dGPUs.

Remember, everyone didn't believe Wccftech and other sites when the problems with 20 nm started to surface. Then it became common wisdom.
Now "everyone" believes that 20 nm won't happen for anyone, even if the CEO of AMD disagrees. The hive mind is a hilarious thing to watch.

 

[raghu78]

Sure, and this is not negated by anything. Most of AMD's products are non-dGPUs. Lisa Su has stated multiple times that they are pushing 20 nm in 2015 for selected GPUs(yes she mentioned graphics specifically), and that only makes sense for their cash cows, i.e. dGPUs.

Remember, everyone didn't believe Wccftech and other sites when the problems with 20 nm started to surface. Then it became common wisdom.
Now "everyone" believes that 20 nm won't happen for anyone, even if the CEO of AMD disagrees. The hive mind is a hilarious thing to watch.

do you have proof that she stated GPUs ? I don't remember her making that kind of product specific statement. if not you are just speculating.

As for 20nm we are at the end of 2014 and other than Apple who is selling product in high volume tell me any other company doing the same. Qualcomm has a modem MDM 9635 shipping in very low volume as does Xilinx with low volume FPGA.

btw you missed the main crux of my argument which is GF has not yet qualified CPI for 20nm / 14nm logic for 2.5D stacking. that happens in H2 2015. So expect 14nm GPUs with 2.5D stacking in mid 2016.

Don't tell that to the FPGA vendors. I think Xilinx was boasting that it had built the world's highest transistor count chip on the 20nm process. And those FPGAs consume serious power.

care to share the die size, transistor count and TDP figures for Xilinx 20nm FPGA. Apple's A8X chip is at 3 billion transistors, 128 sq mm and 5W TDP.

http://www.anandtech.com/show/8716/apple-a8xs-gpu-gxa6850-even-better-than-i-thought

AMD's next gen chip R9 390X is a 8+ billion transistor 28nm behemoth with a die size of 520 - 550 sq mm connected to 4GB of HBM with 512 GB/s bandwidth on a silicon interposer. TDP should be 270W which would explain the reference hybrid cooling design. AMD would want to keep up the good work from R9 295x2 in terms of cooling. These are my rough estimates.

 

[Arachnotronic]

care to share the die size, transistor count and TDP figures for Xilinx 20nm FPGA. Apple's A8X chip is at 3 billion transistors, 128 sq mm and 5W TDP.

20 billion transistors for Xilinx's top 20nm FPGA.

http://www.xilinx.com/publications/archives/xcell/Xcell86.pdf

With the Virtex UltraScale XCVU440, Xilinx is
smashing its own record by offering a
20-nm device with 4.4 million logic cells
(the equivalent to 50 million ASIC gates)
for programming. The device also is by
far the world&#8217;s densest IC, containing
more than 20 billion transistors.

 

[raghu78]

20 billion transistors for Xilinx's top 20nm FPGA.

http://www.xilinx.com/publications/archives/xcell/Xcell86.pdf

http://forums.xilinx.com/t5/Xcell-D...vyweight-FPGA-Champion-The-Virtex/ba-p/391081

"Yesterday, Xilinx announced the world&#8217;s largest FPGA (again!) in conjunction with the Virtex UltraScale and Kintex UltraScale 20nm All Programmable device announcements. The new heavyweight FPGA champion, dubbed the Virtex UltraScale XCVU440 All Programmable 3D IC, weighs in at 4,407,480 logic cells. That&#8217;s more than twice as large as the Virtex-7 2000T 28nm 3D IC, announced in late 2011, which incorporates 1,954,560 logic cells. It&#8217;s also four times larger than any monolithic FPGA available from any other source. Both of these Xilinx 3D ICs take their respective process nodes &#8220;off the Moore&#8217;s-Law chart&#8221; by incorporating multiple large die on a 3D IC. The Virtex-7 2000T 3D IC employs four active die (Super Logic Regions or &#8220;SLRs&#8221; in Xilinx-speak) to achieve the 2M-logic-cell capacity but the Virtex UltraScale VU440 3D IC requires only three die to achieve its world-leading heavyweight capacity of 4.4M logic cells, as shown in this somewhat conceptual illustration:"

thats across 3 dies connected on a silicon interposer. anyway you realize these FPGAs are very low volume and very high priced products with price in tens of thousands of dollars. see slide 7 of pdf below. the last gen 28nm Virtex 7 xc7v-2000t has a price of 16,286.94&#8364;. XCVU440 must be atleast twice that price. moreover these devices draw a lot of power. see slide 13

https://indico.cern.ch/event/283113/session/2/contribution/30/material/slides/2.pdf

the clocks you can achieve with FPGA are nowhere close to what you achieve with a GPU (which is an Application Specific Integrated Circuit or ASIC). FPGA clocks are 500 - 600 Mhz and the more complex the FPGA the lower the max clocks. a high performance GPU can easily clock 1.2+ Ghz which is 2x the clocks of FPGA.

http://www.xilinx.com/support/documentation/data_sheets/ds893-virtex-ultrascale-data-sheet.pdf
http://www.altera.com/literature/wp...ng-peak-floating-point-performance-claims.pdf

Cost is atleast 2 orders of magnitude higher than GPUs so FPGA vendors can afford to eat yield losses as they have very high margins.

 

[Fox5]

I don't see any reason 20nm wouldn't work for large gpus, other than if yields sucked.

GPUs are clocked at low speeds and use low voltages, they're just large and wide. They seem kind of ideal for a low power node. Remember, it was Nvidia with Fermi who tried to up the clock speed on GPUs and that resulted in insane power consumption.

It wouldn't be the first time AMD has beat nvidia to new nodes though. Also, AMD's been heavily advocating automated design tools as of late, so they may be more able to switch to multiple nodes (something they already have to do to go between GF and TSMC) as opposed to nvidia who may be doing more custom design at this point. Nvidia can get an edge on something like Maxwell, but AMD could hit new nodes sooner.

It would be interesting if the GlobalFoundries relationship finally benefits AMD however, and they're able to get capacity on some cutting edge nodes.

 

[NTMBK]

Still expecting AMD to go with HBM, not 20nm.

 

[raghu78]

http://wccftech.com/amd-gpus-28nm-shp-glofo/

wccftech is at it again. their speculation on others speculation will never end. Anyway I am confident of GF 28SHP being the node on which AMD builds its next gen GPUs with HBM. Thats their best (performance) and safest process option (yields).

 

[Bateluer]

As raghu stated, I don't think 20nm is a shoe in for 300 series. It could be the 2nd case that AMD went huge die on 28nm and ran into yields, temperatures or power consumption issues when trying to get clock speeds good enough to hit their performance targets. Either way we slice it, NV is going to bring out 960 next month and AMD will be 4 months behind. By end of March 2015, AMD's 300 series will be 6 months behind similar to 5850/5870 vs. 470/480, at which point every extra month delay would be very costly if they are stuck on the same 280/280X/285/290/290X series. 960 has the potential to wipe out all of their non-290 sales. Are they going to drop 290 to $199 to combat 960/970? We know that people will pay a premium for NV with lower performance so even if a $250 960 is 5-10% slower than a 290, it's game over for the 290. Realistically since 200 series is now old tech, uses more power, it can't be priced similar to NV even with the same performance.

Surely AMD has ran into some issues, maybe even HBM constrained, because when 960 enters, things will go from bad to ugly. Their mobile dGPU market share will get wiped out by GM206 entering laptops.

Your argument falls apart when you take into account that the 290 is already usually cheaper than a 970 while offering similar performance. A 960 would have to be priced under 175 and have at least 3GBs of VRAM to be a viable card.

Not sure why RussianSensation is so doom & gloom about AMD. It flies in the face of every benchmark out there.

 

[raghu78]

Your argument falls apart when you take into account that the 290 is already usually cheaper than a 970 while offering similar performance. A 960 would have to be priced under 175 and have at least 3GBs of VRAM to be a viable card.

Not sure why RussianSensation is so doom & gloom about AMD. It flies in the face of every benchmark out there.

exactly. RS is in fact paranoid when it comes to AMD's impending doom. He has always underestimated AMD. He was one of the most vocal persons before R9 290X launched saying AMD will not be able to even catch up GTX 780. Guess what AMD beat Titan and has aged more gracefully than GTX 780 Ti. You only need to see performance in the major games released in the last 3 - 6 months to know that.

In fact the way Nvidia has treated its Kepler owners I wouldn't be surprised to see a lot more Nvidia enthusiasts go with R9 390X if the rumours are true about it taking the GPU crown.

look at this post here from RS focussing on die sizes and saying R9 390X at 550 sq mm can only be 30% faster than R9 290X . He forgets to see that there are lot of changes happening here. The architectural changes and other chip efficiency improvement techniques combined with the impact of a state of the art High bandwidth memory system (HBM) on a 2.5D silicon interposer which provides much higher bandwidth (> 60% vs R9 290X) at vastly lower power.

http://forums.anandtech.com/showpost.php?p=37013050&postcount=236

With a slightly modified GCN 1.1 architecture R9 290X was 35% faster than R9 280X with 37.5% more shaders. R9 390X has 45% more shaders and will inherit the massive improvements to tesselation performance, ROP performance and memory bandwidth efficiency from Tonga (GCN 1.2).

GCN 2.0 aka R9 390X will in fact improve the core shader architecture too to improve perf/sp, perf/watt and perf/sq mm. He also fails to see that HBM is a game changer as it moves the vast complexity of the on die memory controller logic to the memory stack saving area, transistors and power. HBM cuts power by 65% for a given bandwidth when compared to GDDR5 (memory controller, PHY and GDDR5 chips). see slide 45 of pdf below. The HBM chip is a 5mKGSD (molded known good stacked die) where the base die is logic and then you have 4 stacked RAM chips.

http://www.microarch.org/micro46/files/keynote1.pdf

In a high end GPU like R9 290X with a 512 bit memory bus the memory subsystem can consume anywhere between 35 - 50% of board power. Here is a study of older GPUs Quadro FX 9800 (based on GTX 280 with a 512 bit GDDR3 memory bus) and HD 6990 (with 2 HD 6970 GPUs each with a 256 bit GDDR5 memory bus) where the memory controller & DRAM chips consume 60+ % and 35+% respectively.

http://www.cse.psu.edu/~juz138/files/islped209-zhao.pdf

AMD has other tricks up its sleeve like Integrated voltage regulator (IVR), integrated passive devices (IPD) and adaptive voltage operation which further cut power. These are mentioned in the GF presentation below and also in AMD's power efficiency feature roadmap.

https://www.youtube.com/watch?v=po29B53bpic

IVR and IPD are the logical companions for HBM to cut memory and board power. Adaptive voltage operation is a a feature in the upcoming Carrizo APU. Since AMD's GPU and APU share a lot of the work on efficiency improvements its logical to see this tech make its way to next gen GPUs. Carrizo is also most likely to feature an IVR to improve power efficiency.

http://images.anandtech.com/doci/8742/Voltage Adaptive.png
http://images.anandtech.com/doci/8742/Carrizo Efficiency.png

Lastly the GF 28SHP process is a significantly better process than TSMC 28HP at which R9 290X was built. The vastly lower leakage power figures for Beema (GF 28SHP) compared to Kabini (TSMC 28HP) confirm the same.

http://www.anandtech.com/show/7974/...hitecture-a10-micro-6700t-performance-preview

"AMD claims a 19% reduction in core leakage/static current for Puma+ compared to Jaguar at 1.2V, and a 38% reduction for the GPU. The drop in leakage directly contributes to a substantially lower power profile for Beema and Mullins."

Leakage power amounts to 30 - 35% of a chip's TDP.

http://www.realworldtech.com/intel-45nm-hkmg/4/

"To give an example of the benefits that the 45nm process could bring to a microprocessor, we will examine a hypothetical scenario. Tulsa, a 65nm implementation of the Pentium 4 microarchitecture, has plenty of available data from ISSCC and Hot Chips. Tulsa dissipates 150W at 3.4GHz, with 16MB of shared L3 cache. Of that 150W TDP, roughly 45W is from leakage power, while the remainder is active power dissipation [2]. "

For a 250W GPU a 38% reduction in leakage power (which when assumed as 30% of chip TDP = 250 * 0.3 * 0.38 = 28.5w) means a 28W power reduction. Anyway I will be glad to see him be proven wrong once again as AMD achieves the huge 60 - 65% perf jump of R9 390X using high performance 28nm process (GF 28SHP) with a die size <= 550 sq mm. R9 390X is going to be an awesome performer.

 

[RussianSensation]

Your argument falls apart when you take into account that the 290 is already usually cheaper than a 970 while offering similar performance. A 960 would have to be priced under 175 and have at least 3GBs of VRAM to be a viable card.

Not sure why RussianSensation is so doom & gloom about AMD. It flies in the face of every benchmark out there.

Because the average gamer who buys NV isn't like you or I that will read countless reviews to find a cool and quiet after-market 290. If 960 is within 10% of 290 in performance with 120-130W power usage, this gamer will buy it over the 290 at $249. 290 series is plagued by high noise levels and high temperatures image. Also, most reviewers will use a clock throttling 290 and possibly against after-market 960, which means automatically 7-10% of lost performance compared to after-market 290 in professional reviews and 960 looks even better due to much higher boost achieved by after-market versions.

@ Raghu,

You misinterpreted my post:

"If 390X is 550mm2, that's a 25.6% increase in die size over 290X which may get you 30-35% increase over 290X, but nowhere near 50-65%. Also, 290X's power usage went even higher than the 7970Ghz. That means for AMD to bring 50-65% increase in performance at 290X's power usage, the architecture has to be significantly redesigned compared to Hawaii."

Nowhere do I state that the final 390X will end up at just 30% faster. I am saying that if it's > 30%, significant changes to the architecture would take place which is a possible reason for such a lengthy delay vs. 970/980 as it takes a lot of time to make major changes to architectures. Secondly, you make a comparison of 290X vs. 7970Ghz but ignore that power usage skyrocketed from 238-243W to 280-290W. You cannot just magically add 45% more SPs/TMUs on 28nm blow way past 290X's power usage with HBM alone. Architecture and dynamic voltage switching must be far more advanced than what's in Hawaii for you to get 60-65% more power over the 290X. My post clearly reflected all of that and yet you assumed I said AMD will never achieve that. I never said that it's impossible; what I said is how it will require dramatic changes which are probably very hard to implement hence 300 series is nowhere to be found.

Regarding 290X not catching up to 780Ti, I could never predict that NV would throw Kepler under the bus. I've never seen NV doing that to their last gen flagship cards 1 year from their release. When 280X is coming close to 780/Titan in some cases, you probably agree AMD is not 100% responsible for such outcomes.

 

[raghu78]

Because the average gamer who buys NV isn't like you or I that will read countless reviews to find a cool and quiet after-market 290. If 960 is within 10% of 290 in performance with 120-130W power usage, this gamer will buy it over the 290 at $249.

@ Raghu,

You misinterpreted my post:

"If 390X is 550mm2, that's a 25.6% increase in die size over 290X which may get you 30-35% increase over 290X, but nowhere near 50-65%. Also, 290X's power usage went even higher than the 7970Ghz. That means for AMD to bring 50-65% increase in performance at 290X's power usage, the architecture has to be significantly redesigned compared to Hawaii."

Nowhere do I state that the final 390X will end up at just 30% faster. I am saying that if it's > 30%, significant changes to the architecture would take place which is a possible reason for such a lengthy delay vs. 970/980 as it takes a lot of time to make major changes to architectures.

The timing of the next gen AMD GPUs products is constrained by primarily GF and Hynix which need to get 2.5D packaging and HBM ready for volume production. You realize that GF has stated that they are production ready for 2.5D stacking in Q4 2014. Hynix is in production of stacked DRAM with customer level qualification samples achieved in Sep 2014 and volume production in Q1 2015. (slide 3 from sk hynix presentation below)

http://www.memcon.com/pdfs/proceedings2014/NET104.pdf
https://www.scribd.com/doc/248650452/GlobalFoundries-2014-US-Tech-Seminar-Proceeding-book-pdf
http://www.kitguru.net/components/g...ogue-set-to-start-mass-production-in-q1-2015/

The production process for 2.5D stacking products is complex and lengthy and happens both at foundry (GF) and at OSAT partners (Amkor) (see slide 4 of Advanced Packaging Update in the GF presentation above) we can be fairly sure that GF is producing the R9 390X and R9 380X chips now for back end integration at OSAT partners in Q1 2015.

2.5D stacking is a disruptive and revolutionary technology and bringing it to market in high volume with good yields and at affordable prices is a massive challenge. So more than anything these are the factors which lead to a late Q1 2016 release. But rest assured once the chips are launched the time to market advantage which AMD has will be an irrefutable advantage for atleast 18 months.

I am looking forward to notebook versions of Fiji and Bermuda with lower clocks, lower voltages and aggressive power binning. The massive bandwidth advantage of HBM in power constrained form factors will be clearly evident. HBM1 clocks are in the 800 Mhz - 1.2 Ghz range. Fore notebooks even when running at 800 Mhz the bandwidth is a stunning 409.6 GB/s for a 4Hi 4 stack config with 4GB HBM. The current GTX 980M runs GDDR5 memory at 2500 Mhz x 2 = 5 Ghz for 160 GB/s bandwidth. Even if Nvidia release a full version of GM204 with 3000 Mhz GDDR5 which will increase power significantly the bandwidth will be a paltry 192 GB/s. Now such a chip against the mobile version of Fiji XT with 400+ GB/s will be a one sided contest especially for 4k gaming notebooks. Two mobile Fiji XT with XDMA CF will scale brilliantly and be pretty much dominant. AMD can get mobile Bermuda XT into a 120W TDP and then its pretty much a gut wrenching knock out for GM204.

Remember Nvidia cannot fit GM200 in MxM form factor as its 384 bit memory bus cannot fit in MxM which are limited to GPUs with 256 bit memory bus. The only option is to disable 2 memory controllers of GM200 and try to run the full GPU with 256 bit memory bus. But that will be a hopelessly bandwidth bottlenecked chip.

http://en.wikipedia.org/wiki/Mobile_PCI_Express_Module#2nd_generation_configurations_.28MXM_3.29
http://www.mxm-sig.org/file.cfm?doc=36A88883-DFFA-E3E6-16C301BCF5633572

 

[RussianSensation]

I doubt AMD will use Bermuda XT in mobile at 120W. They will most likely use Fiji, just like Hawaii never came out as a 120W binned laptop version. It also makes little sense to give a 120W 300 series chip 400GB/sec memory bandwith. It won't have the GPU power to take advantage of so much bandwidth.

You say AMD will have an 18 months lead but that would mean 380X > 980 and 390x > GM200. We don't know that. Furthermore, in the deaktop NV can lower 980 to $399 by August 2015, release GM200 780 successor at $499-550 and 780Ti successor at $699. That doesn't even account for 970Ti and a faster closed 980 with 8GB of VRAM.

On the laptop side, by May-July 2015, NV could introduce 975M and 985M, with more CUDA cores/faster clocks. Using 970M/980M/970/980 as a benchmark for your May 2015 or later 300 series doesn't take into account that NV won't stay still and that prices will remain at today's levels. Remember one year after 680's launch, NV easily bumped performance 7-10% and lowered the price from $499 to $399 at the same time.

Every month AMD is late means more time for NV to prepare refreshes of GM204 and milk existing GM204 so much that the higher profit margins since Sept 2014 to May 2015 will more than allow for NV to drop prices on GM204 to defend market share. For AMD to make major leaps in 2015-2016, R9 300 series has to be spectacular, maybe their best ever product since 9700 Pro.

Since AMD don't wait AMD cards 6-9 months late like NV users would, every month there is a major risk of AMD users buying GM204 products and only looking for a new GPU upgrade in early 2017-2018 for the next round. NV is probably getting 90% of $200+ mobile dGPU sales right now and 90% of $400+ desktop sales, if not more. This is going to continue until 300 series launches. Ouch.

 

[tviceman]

Russian why you no respond to my private messages?

 

[Rezist]

HBM sounds cool and all but at what costs? I understand AMD moving everything to GF, just due to the WSA but that will bite them when TSMC has a lead over them and GF doesn't want to spend any money on R&D to be competitive, knowing they have AMD by the balls.