In-Depth Analysis: Intel Roadmap Execution since 2014

[witeken]

In-Depth Analysis: Intel Roadmap Execution since 2014

Analysis. Brian Krzanich became CEO on May 16, 2013. About how much of these delays have been caused by his (lack of?) proficiency, and other managerial shuffling going on, I shall not speculate. Instead, three other root causes can be found.

First is the aggressive mobile project that P. Otellini had started, which was eventually watered down (starting end-’14, whether by BK himself or forced by the Board of Directors is open to speculation) in favor of putting more resources into the vastly more profitable data center, which culminated in the major 2016 restructuring with the almost complete abandoning of the mobile market (save for the partnerships & modem).

Secondly, the (relatively) low 14nm yields (and consecutively the 10nm ones as well) have had a cascading effect on the whole launch schedule in general. Given the (somewhat lack of) competitive environment, it can be expected that management chooses to favor margins instead of (aggressively) moving down to a less healthy node anyway.

Lastly, the decision to move the data center to the leading edge first has likely caused a serious (most recent) reshuffle. For instance Intel probably had to insert a fourth 14nm cycle (Coffee Lake) into the consumer product line to be sure to keep a yearly product release cadence (they might have also done that at 10nm, but likely decided to stay on 14nm longer given the status of 10nm yield), as the move from 10nm to 7nm would have been prolonged by the strategy shift to data center. Except for general delays, the biggest victim of this decision was the cancellation of Cannonlake-EP in favor of an aggressive Icelake-EP schedule. The 3D XPoint delay seems to be mainly because of technical R&D reasons. The only delay I have identified to be because of valid reasons instead of development issues, is silicon photonics*.

Whether Intel can improve the shady execution track record they have assimilated in the last four years, remains to be seen. However, several improvements can be noted. As the shift of focus and reorganization starts materializing, a less volatile roadmap should appear. As Intel aligns itself to the not aggressive 3 year Moore’s Law cadence, less delays should appear hand in hand with the first new fundamental lithography shift in more than a decade; EUV.

I think long-term Intel’s commitments remain relatively unchanged (i.e. pursue Moore’s Law as IDM and yearly product cadence; data center with cloud, network, HPC, AI & silicon photonics; IoT with autonomous cars; 3D NAND; 3D XPoint; FPGA; foundry), but needless to say delays cause a less compelling product offering than originally planned.

As a final remark, I’m not sure any other tech company has such a visible roadmap. I have not tried to compare Intel to other companies. But I’ll just say that TSMC once promised 20nm in early 2013 and 10nm (back then still called 14nm) in early 2015. Also, before mid-'16, people were very impatient about the lack of big new GPU families, with Nvidia and AMD sticking for four or more years to 28nm node. Needless to say their 14/16nm products have also suffered major (internal) delays.

Product delivery dates Intel promised but did not meet (42 products)

• Merrifield: Q4’13 became Q1’14
  
• 14nm production: HVM in Q1’14 became Q3’14 (with yield parity in H2’16)
  
• Broadwell-Y: Q2’14 became Q4’14
  
• Moorefield: Q2’14 became Q4’14 (Only major design win, ZenPhone 2 w/XMM7260, launched in Q1’15.)
  
• Cherry Trail: Q3’14 (mid’14) became Q4’14 became Q2’15
  
• Broadwell-U: Q3’14 became Q1’15
  
• SoFIA 3G (28nm): Q4’14 became Q2’15
  
• Broadwell-C: Q1’15 became Q2’15 (mid’15)
  
• Silicon photonics*: Q1’15 became Q3’16
  
• Skylake-S/-H: Q2’15 became Q3’15 became Q4’15
  
• Willow Trail (aka Morganfield SoC for tablets): Q2’15 became canceled
  
• Morganfield SoC (2.7GHz 14nm QC Goldmont w/Gen9): Q2’15 became canceled (Q4’15 for Android)
  
• Broxton: Q3’15 (mid-‘15) became Q4’16 became canceled
  
• XMM 7360: Q3’15 became Q4’15. “First commercial devices”: H2’15 became Q3’16
  
• SoFIA LTE (28nm w/Spreadtrum): Q3’15 (mid’15) became Q2’16 became canceled
  
• Skylake-U: Q3’15 (mid’15) became Q4’15
  
• Broadwell-E HEDT: Q4’15 became Q1’16 became Q2’16 (mid-’16)
  
• Apollo Lake: Q4’15 became Q3’16
  
• Xeon Phi Knights Landing: H2’15 became Q2’16
  
• Associated with KL also Omni-Path Architecture and Hybrid Memory Cube: H2’15 became Q2’16
  
• 10nm: Q4’15 became Q3’17
  
• 3D NAND: H2’15 became Q3’16
  
• 3D XPoint: 2016 became Q4’16 became mid-‘17
  
• Riverton (low-end 14nm Airmont with 4G modem): Q1’16 became canceled
  
• Broadwell-EP/EX: Q1’16 became Q3’16 (before that planned for 2015
  
• Kaby Lake-U/Y: Q3’16 became Q4’16
  
• Kaby Lake-S/-H: Q4’16 became Q1’17
  
• Skylake-S GT4+4e: Q4’16 became canceled (65W; 35W version considered but not greenlit)
  
• Cannonlake-U/-Y: H2’16 became Q3’17 became Q4’17 became Q1’18 (volume)
  
• Cannonlake-H/-S: 2016 became canceled
  
• SoFIA 14nm w/Spreadtrum: H2’16 became Q3’17
  
• SoFIA LTE2 14nm: Q4’16 became Q3’17 became canceled, along with it the…
  
• Integrated 14nm modem (XMM 7560): Q4’16 became Q3’18 [per BK, no clear roadmap found]
  
• Skylake-EP/EX (Purley): Q2’17 became Q3’17 (mid-’17) (before that planned for H2’16)
  
• Skylake-X: Q2’17 became Q3’17
  
• Apache Pass (3D XPoint DIMM for data center along with Purley): Q2’17 became cancelled
  
• Kaby Lake-X: Q2’17 became Q3’17
  
• Gemini Lake: Q4’17 will become ???
  
• Knights Hill: Never formally disclosed, but possibly H2’17 became H1’19
  
• Omni-Path 2nd gen: Along with Knights Hill
  
• Coffee Lake-H/-U: Q2’18 became H2’18 [confirmation needed]
  
• Cannonlake-EP/EX: 2018 became canceled
  
• Icelake: 2018 per PAO, possibly delayed until 2019?
  
• Tigerlake: 2019 per PAO, possibly delayed until 2020?

In order of biggest delay (excluding canceled products)
3Q: Cherry Trail, 3D NAND, BDW-E, Apollo Lake, KNL/OPA/HMC, ~3DXP, Spreadtrum 14nm Atom
5Q: Broxton, ~Knights Hill/OPA2
6Q: Silicon photonics
7Q: 10nm (Cannonlake), 14nm modem (?)
11Q: 14nm yield parity (at same time in development)

Products released on schedule (~3 products)
Braswell: Q1’15 (one might argue delay until Q2’15)
SoFIA 3G-R (28nm quadcore w/Rockchip): Q2’15 (Delivered in Chinese market, “selling great”--BK.)
Arria 10 w/BDW-EP PCIe deep learning inference card: 2017 (no surprise with a 12 months launch window)
One might also argue SKL-U, or other products that slipped only 1 quarter.

Products announced/leaked not yet released but no known delay

• Lake Crest: H2’17
  
• 3D NAND Gen2: Q3’17
  
• Knights Mill: Q4’17 [4x faster at deep learning]
  
• Coffee Lake-S: Q2’18
  
• Icelake-EP: Q1’19 [10nm++]
  
• Sappire Rapid (Xeon Tinsley): Q3’20 (mid-‘20) [7nm]
  
• KNP (7nm Xeon Phi): Q3’20 [7nm]
  
• Knights Crest: [Manufactured in Intel fabs by 2020]
  
• Rumor: Tock on 7nm: new architecture without full BC: http://www.bitsandchips.it/english/52-english-news/7854-rumor-even-intel-is-studying-a-new-x86-uarch

No date available:

• Coffee Lake-X

Notes and remarks

• All dates have been taken directly from both public and leaked Intel confidential roadmaps. This document should be fairly accurate.
  
• For some products like Xeon-D, Atom for microservers and other minor products, no roadmap was available or I did not consider interesting, so those have been omitted.
• I'll also say here explicitly that Intel is not the only company that suffers delays. However, some products have been executed much worse than others, and some products have simply not even been very exciting as well in terms of performance increase (Skylake) or competitiveness (GPU and modem roadmap).

• ExtremeTech notes that the status of SoFIA 3G is announced, not launched, interestingly.
  
• ExtremeTech also notes that, according to Intel, their partnership with Rockchip (as of April 29, 2016) has not been terminated. Intel formed the partnership in 2014 to get a SoFIA quadcore into the Chinese market, as Intel itself was only working on a dualcore SoFIA (but Chinese market don’t like low core counts). The partnership with Spreadtrum started at MWC 2014. Intel invested $1.5B in Tsinghua Unigroup, although Spreadtrum is under no obligation to use IA exclusively. Spreadtrum is also working on 10nm SoCs (although it’s not the only mobile company, with LG being the other). (https://www.extremetech.com/computing/227523-intel-kills-upcoming-smartphone-and-tablet-hardware, http://www.eetimes.com/document.asp?doc_id=1324100, http://www.eetimes.com/document.asp?doc_id=1326692, http://www.eetimes.com/document.asp?doc_id=1331424, http://www.spreadtrum.com/en/show_news.html?id=fe766282-e9cc-4ffd-b211-e3d19675d3f7, http://www.anandtech.com/show/10288/intel-broxton-sofia-smartphone-socs-cancelled)
  
• The cancellation of SoFIA “3GX” (Intel’s words): This is weird. I couldn’t find which product that is, despite having been reported by numerous media as being cancelled. I suppose this means all SoFIA 3G products were discontinued. Cancellation of SoFIA 4G 28nm, SoFIA 4G 14nm and Broxton were confirmed during the announcement of the major 2016-2017 restructuring following Q1’16 earnings. Likely Murthy had a large role in this move, who joined Intel from Qualcomm in November 2015 following Qualcomm’s decision to move to a 1-person leadership for the QCT group, thus ditching him. (http://telecom.economictimes.indiatimes.com/news/devices/intel-hires-qualcomm-executive-murthy-renduchintala-to-help-expand-ailing-pc-business/49887702)
  
• As reported by Ashraf Eassa, this does not, however, mean that Intel has definitely shied away from the mobile market – as is also evidenced by the continued Rockchip and Spreadtrum partnerships. He notes that Murthy found the products not competitive. (https://www.fool.com/investing/2016/12/06/intel-corporations-murthy-renduchintala-hints-at-n.aspx)
  
• Despite Skylake’s delays, the -U and -Y SKUs followed remarkably close to the BDW-U and -Y series, within less than a year.
  
• The XMM roadmap has been largely omitted due to it never having had a solid roadmap with verifiable milestones, given the lack of major design wins. Intel’s statement about the 4G roadmap has always been to get a yearly cadence. This does not seem to have materialized completely. Per Intel Ark, the XMM 7160 launched in Q1’13, the XMM 7260 in Q3’14, the XMM 7360 in Q4’15. The XMM 7480 has not even been placed on Ark yet, but was verified in Q3’16 and has started sampling. Analyzing this further, this lack of strict yearly cadence has put Intel continually further behind Qualcomm. The first commercial devices with XMM 7360 were supposed to come in H2’15, suggesting that with a yearly cadence the iPhone 7 should have gotten an XMM 7480 (although it was from the start rumored to ship with XMM 7360, probably to take no risks). Whether this slippery roadmap is because of an early shifted focus to 5G is not sure, but possible.
  
• Intel said (MWC’17) XMM 7560 would start sampling in H1’17 with productions “soon afterward”. However, at their Q3’16 earnings call, Brian Krzanich strongly indicated volume for this product should be expect in H2’18, according to him in line with the yearly cadence.
  
• SoFIA 4G was meant to ship with a low-end, budget version of XMM 7260.
  
• Many more mobile projects were in development or planning stage than this document mentions. See source below for the full Intel mid-2013 mobile roadmap.
  
• Riverton likely was already canceled from the roadmap in 2013 in favor of the new SoFIA roadmap announced at Investor Meeting November 2013.
  
• For 3D XPoint, Intel did not provide a more specific date than 2016 launch, with sampling starting in 2015. The delay could be anywhere from 6 months to more than a year.
  
• *The silicon photonics delay was (completely?) due to decision to transition to 300mm wafers. In 2013, Intel promised production would start in 2014, but I have placed it in Q1’15 since volume was only expected to grow to a meaningful level in 2015, which did not materialize.

Select sources
http://wccftech.com/intel-atom-soc-roadmap-tablets-smartphones-leaked-14nm-airmont-goldmont-based-chips-detailed/
http://cdn.wccftech.com/wp-content/uploads/2015/11/Intel-2016-Roadmap-635x357.jpg
http://www.cnx-software.com/wp-content/uploads/2016/09/Intel-Roadmap-2017-2018.jpg
http://cdn.wccftech.com/wp-content/uploads/2016/06/Intel-Skylake-X-and-Kaby-Lake-X-Leaked-Roadmap.jpg
https://www.overclockers.at/attachment.php?attachmentid=195098&thumb=750
http://cdn.wccftech.com/wp-content/uploads/2015/11/Intel-2016-Roadmap-635x357.jpg
http://fs5.directupload.net/images/161229/9gxhy79y.png
http://wccftech.com/intel-14nm-skylake-ep-10nm-cannonlake-ep-supported-purley-platform-160w-tdp-48-pcie-lanes-6-channel-ddr4/
http://i.imgur.com/jvQaVLH.png

Update October 2017
At the end of the year, how has Intel done.

• Several IoT products killed
• Kaby Lake launched on schedule
• Skylake-X on or ahead of schedule
• Coffee Lake is launching on or ahead of schedule
• Cannonlake is still Q1, but volume on 10nm (ICL-U/Y, ICL-S uncertain might become Q4'18
• Gemini lake seems on schedule for Q4'17
• Knights Hill probably canceled, my speculation is Xeon Phi will be replaced by a better EMIB-based many core, else I don't understand what Intel will position against Nvidia
• Spreadtrum deal for 14nm (and 10nm) SoCs: does anyone know?
• Kngihts Mill: will probably launch with November Top500
• Lake Crest: hopefully it will also be launched then
• FPGAs...
• 3D XP...
• 3DNAND: 64 layer launched on schedule

 

[NTMBK]

"3 year Moore’s Law cadence" is an oxymoron.

 

[IntelUser2000]

Apache Pass isn't cancelled. It's just delayed until next Purley refresh likely in 2018. The whole Optane deal hinges on the DIMM part because compared to it, others are essentially useless. The reason they were able to converge EX and EP is likely thanks to Optane. The EX chips always had expensive and slow buffer chips that allowed extended memory capacity. Optane allows the extra capacity without cost and power increase of the buffer chip.

Of course delays mean it won't be good as expected. At least the density advantage compared to DRAM will decrease.

The Optane for Q2 is only for caching. The SSDs are even later.

Purley is likely coming before SKL-X parts. It was with Broadwell, and they deem server more important, no reason to change.

 

[Dresdenboy]

Thanks for this interesting overview! I'll use it as an inspiration for some aspects of a game.

 

[dfk7677]

So Coffelake-S (Desktop) will come out in 2Q18? And no Cannonlake for Desktop?

 

[tamz_msc]

Apparently, Charlie@Semiaccurate says that 3D Xpoint is underwhelming and Intel had a press event on 15th March to paint a rosy picture of it.

 

[IntelUser2000]

Apparently, Charlie@Semiaccurate says that 3D Xpoint is underwhelming and Intel had a press event on 15th March to paint a rosy picture of it.

3D XPoint is underwhelming for the soon to be released products, but that's not looking at the whole picture.

The article is mostly complaining about the cache and the SSD, and even then shows bias(At some point Charlie became like this. Whether it's due to Intel refusing to give him insider secrets like they did back in the Conroe days or other reasons I don't know). The cache being successful is mostly how they will price the thing, and the SSD as well. The latter beats the NAND devices in all the important metrics, so people who need the performance will go for it provided it's competitively priced.

He points about the endurance being only 3-10x for the SSD, and it's true it falls far short. But there's no reason they won't put low quality, cheaper chips for the SSD. Every semiconductor has cut down chips, why not this one?

It especially makes sense since the delay is likely due to yield issues. But there's no denying the technology has potential. The DRAM versions will use the highest quality chips with best endurance, performance, and be most expensive.

 

[tamz_msc]

Obvious bias of Charlie aside, he does make an interesting point - where does Micron feature in all of this? After all it was co-developed with Micron.

 

[raghu78]

Intel's execution between 2014 and 2017 will determine the next 3-4 years of their product roadmap. I do think that Intel manuacturing has lost their leadership once and for all. TSMC 7nm will be extremely competitive with Intel 10nm and the furious pace at which Apple and TSMC are driving process nodes, it is not going to be easy for Intel to keep up. If the rumours of Samsung introducing Gate all around FET in 2018 with their 7nm process turn out true then it will be a mighty embarassment for Intel which led these foundries by 1-2 process node generations (2-4 years) and was a leader in breakthrough innovations in transistor device engineering and process technology for decades.

 

[scannall]

The explosion of mobile has given the foundries a large infusion of cash, to push R & D. And has allowed them to close the gap with Intel. Intel may still be 'Chipzilla'. And their process is still a *bit* better overall. But it's a precarious lead at best. And theirs to lose. ARM is also a problem for Intel. Mobile OEM's just looked at the history of PC OEM's fighting for scraps, while Intel and Microsoft took the bulk of the cash. Clearly, putting x86 in your mobile device isn't something OEM's were willing to do. Not even wrapping their x86 SOC's in $20 bills.

Looks like AMD is finally competitive again, which is a great thing. And my next build will be a Ryzen. Puts more pressure on Intel though.

Is Intel 'DOOMED!'? No, of course not. It makes for fun tech drama, and forum trolling fodder. But Intel is going to have to get its act together. They can't afford to keep making so many missteps, as there are now competitors ready to pounce.

 

[trivik12]

Pat Gelsinger mentions a new product "Aero Lake".

And then Meteor Lake, Lunar Lake, Aerolake, Panther Lake, we have a very robust roadmap on the client side. So all of that taken together, AI is a workload and we are going to make it available everywhere and that’s what our overall strategy’s about. Compete at the high end, deliver it in Xeon, but make it available at the client and edge.