Intel Investor Meeting 2017: February 9

[witeken]

Intel Investor Meeting 2017: February 9

Webcast + PDFs

This will be my third Investor Meeting coverage at AT here (in November I also covered the 1st Intel AI day), the last one was held in November 2015, but the 2016 got pushed into 2017 due to the change of CFO because Stacy Smith got promoted.

Intel will normally release a webcast (livestream) for those who would want to watch it live. For the other people, I will post summaries of the presentations with the relevant key announcements.

Normally, the CEO will kick-off with a presentation about what the company's been up to high-level. There will also be presentations from the (new) CFO, the executives from the main groups like PC and data center, and there will be 1-2 Q&As maybe. And I hope we'll get a presentation about the status of Moore's Law at Intel. Since Bill Holt has retired last year, I'm not sure who's going to do it.

The Moore's Law presentation will in any case be the most important one. With other companies' "10nm" already in production and with TSMC's 7nm to follow next year for iPhone 8s, the question is how much lead Intel is going to retain after the delayed 10nm after the delayed 14nm, certainly if 7nm only ramps in 2021. So hopefully we'll hear a bit more about the health of their 10nm process.

In the meantime, feel free to read up on what Intel's been up to, underneath.

Intel
Because it turns out mobile isn't all the lucrative except for a few people, Intel has decreased their investments substantially. Early 2016 shortly after the Altera acquisition, Intel did a big reorganization to further reduce spending for mobile and for the declining PC market. Despite the reorg, however, R&D spending for 2016 in fact still increased by more than $0.5B to almost $13B.

The reason spending has increased is because of the Altera acquisition and others. Intel is currently doing a lot of investments in autonomous vehicles and IoT in general, in artificial intelligence, in memory, and as always, Moore's Law.

Intel's most notable acquisitions last year were the Nervana AI one, which will go up against Nvidia's GPUs, and Soft Machines, which has created the VISC ISA.

For 2017, R&D will probably be about flat, but Capex will surge from less than $8B in 2015 and $10.5B last year to $12.5B this year, driven by 10nm, by 7nm pilot lines, by 3D NAND and by 3DXP. (Although before 2015 it was at $10-11B for quite a few years, so it's not unexceptionally.)

Lastly, financially, Intel has had a good 2016. Despite a weak 1st half of the year, Intel's PC (~$30-32B) revenue actually grew in 2016 -- surprisingly, also in part due to about $0.5B revenues from Apple modem win. The data center, which is almost $18B now and which Intel since at least 2013 has forecasted to grow 15% annually, grew only 8%, after about 11% last year. IoT is growing steadily at about 15% per year too, and nearing $3B. NAND (~$3B) had a bad year, but because of the 3D NAND transition, Q4 was 25% higher YoY.

In early 2017, the McAffee (aka Intel Security, about $2B) will be shed from the company.

All in all, Intel grew revenues from $55B to almost $60B in 2016, but they forecast flat revenue for '17.

Key products

• Core architecture and related IP for PC (Wifi, modem, Atom, I/O, graphics)
• Xeon & Xeon Phi for data center
• More data center IP: Omni-Path HPC interconnect, silicon photonics, rack-scale architecture
• 3D NAND JV with Micron
• 3D XPoint
• AI investments
• Autonomous cars investments
• IoT

Key 2017 products

• 10nm Cannonlake-U/Y (ramp in '18, though, 3yrs after BDW)
• Skylake-X & Kaby Lake-X
• Coffee Lake?
• Purley: 32-core Skylake Xeon
• 3D NAND Gen2
• 3D XPoint (only as SSDs probably)
• Stratix 10 14nm FPGA with Intel's 2.5D (EMIB) interconnect
• Knights Mill: basically Knights Landing with half-precision support for AI
• Lake Crest and Knights Crest for AI (Nervana)
• XMM 7480 modem (28nm presumably) and 5G modem sampling

Key 2018 products
Q1
800P Optane
Skylake-D (embedded Skylake-SP/Lewisburg PCH up to 18C)
64L SSD for data center
Stratix 10 TX (58G transceivers)
Q2
Kaby Lake-G with Radeon Vega and Kaby Lake-U i3 (w/Turbo)
CFL-H (up to i9), CFL-U (with GT3e) and CFL-S with 300-series chipset
Arria 10 Programmable Acceleration Card (PAC) for data center

3D NAND Gen 2 w/QLC
64L QLC 3D NAND
XMM 7560
400G silicon photonics
CFL 8-core with Z390
Whiskey Lake
Cascade Lake-SP/X
3D XPoint DIMMs
Spring Crest
EyeQ4
Summaries

• $7B Fab 42 announcement for leading edge 7nm production in next decade.
• Xeon "shatters" benchmarks

• BK talk: strategy overview
• Murthy Renduchintala: Leading the Data Revolution
• Diane Bryant: Data Center
• Doug Davis: IoT and ADAS
• Bob Swan: Financials
• Q&A

Alright, that's a wrap.

 

[masimilianzo]

I just hope they don't go the nonsense-naming route of nodes like Samsung/TSMC

Inviato dal mio ONEPLUS A3003 utilizzando Tapatalk

 

[witeken]

* I've already added 2 news bullets to the summary.


Brian Krzanich: Strategy Overview

TL;DR:

*Great 2016 business performance + restructuring

*Intel = data company (hmm, few years ago it was the smart+connected=best_at_Intel phrase, now it’s data virtuous cycle)

*A lot of potential for revenue (TAM = total addressable market) growth for the company in the future --> investing

*BK has explained the Intel philosophy in a very nice story of what they’re doing. but how their new investments will play out in the next decade and in terms of execution and revenue wise, remains to be seen of course, but good talk

• PC: units down, ASP up due to segmentation
• Data center: cloud mid-25% growth, networking fastest growing segment (preparing 5G): MUST have software-defined network (instead of proprietary hardware), which is what Intel does
• IoT: specific focus on retail, industrial, security (cameras)
• Altera/PSG: phenomenal, seamless job of (tightly) integrating Altera: growing faster than market (share growth) + 1st 14nm shipments
• Memory (NSG): Dalian fab started, Optane started shipping
• New Technology Group (NTG): being part of those (RealSense, drones, robotics, VR Alloy): interesting because they generate lots of data
• Restructuring / leadership changes: 2016 biggest year of transition, bringing outside perspective
• Sales force: aligning sales teams, tighter link between sales force and business units: e.g. one specifically focused on IoT, job to win design win, one specifically for data center
• Cloud of today = people, cloud of tomorrow = machines, much more data (cars, smart factories, replay tech) -> needs cloud, apply analytics, push data back out
• Intel monopoly? Intel naïve TAM = 32B PC + 13B DCG = 45B
• Intel TAM view: 30B PC + 65B data center + 55B non-volatile + 40B modem + 30B IoT = 220B (note: no, they do not think they can get 100% MSS, I think )
• Intel = rack company + optimize between racks and across the data center (photonics, 3DXP, FPGA, Xeon, Xeon Phi, Nervana, Omni Path, rack-scale)
• 3D NAND = leadership, because of engineering decisions + 3DXP: pushes limits of data very near CPU for high performance (Quote: “every single gamer will want to have 3D XPoint”)
• 5G: end-to-end = big advantage
• IoT: big word, for Intel focused on retail, industrial, security + ADAS (auto. driving): everything comes together there (“only company … deliver end-to-end solutions”)
• Virtuous cycle strategy: (1) Experiences: cloud infrastructure + AI (“unique end-to-end assets”) + networking (fastest growing segment in DCG with low market share)
• (2) Connectivity: Memory + FPGA (lots of products coming to market + looking at 10nm) + 5G
• (3) Edge devices that require lotta data
• Partnering, not creating new things ((un)like the 4G Intel standard), smart M&A (Nervana, Movidius), not doing things that they’re unfamiliar with-> everything = compute related with silicon and software leadership
• Need to “flawlessly execute” the new products like 3DXP

Next up: Murthy!

 

[liahos1]

murthy heavy breathing.

 

[liahos1]

murthy talking about farm to table artisanal bread now.

 

[liahos1]

a full day on intel manufacturing in 1h of this year. Fantastic!!

 

[witeken]

a full day on intel manufacturing in 1h of this year. Fantastic!!

What do you mean?

 

[liahos1]

What do you mean?

Murthy mentioned that Stacey and someone else (missed the name) are going to hold a specific intel manufacturing investor event in 1h of this year. Which is nice because its a whole event dedicated to the process technology vs a segment during today's analyst day.

 

[witeken]

Murthy Renduchintala: Leading the Data Revolution

TL;DR:

*Explains why joined Intel: wanna be part of an IDM, work for number 1 semi-industry, etc. (missed this part)

*Focus on growth segments, costs were cut in 2016

*Shows epic graph of Moore’s Law logic cell area! So far from 2013 onward Intel has every year put a ‘controversial’ graph about Moore’s Law in the presentation, and this year is no different. (see link at bottom of post) looking forward to the debate this will cause

*IDM = good

*First 4G 14nm product in 2018

• Has broad role in the company, has to align process dev and roadmap with the product dev
• Client group: operating profit up 30% YoY, by cutting spending and targeting R&D at growth products to drive up ASP
• Example: E-Sports -> Broadwell-E, 10-core product with ASP over 1k dollar (LOLOLOL, Zen gonna eat this biz?)
• Adjacent businesses do targeted R&D at growth segments (>2B rev): progress in modem (and roadmap), gateway business (Lantic (ph) acquisition), wifi business market share growth, thunderbolt
• ASP is highest since 2011 in PC
• Process tech: “profoundly proud to work for company that .. defines Moore’s Law”.
• Effective logic cell area graph: shows graph from 2007-2018 and shows that Intel is still on same cadence even with Tick-Tock-Tock AND 3 year ahead of other (all on the graph!)
• Announce 8th gen core 10nm, with 15% higher SysMark perf
• IDM: Develop process tech and product simultaneously (customized IP and transistors for specific products). Analogy of baker from farm to bread :’) (look up yourself, lol)
• IDM example: granularity of different transistors for CPU vs. GPU on same die
• Repeating the virtuous growth cycle (see BK) (“interconnectedness of everything”, for real): 5G, VR (with Thunderbolt and Wigig), ADAS, AI
• Going from Core i3 to Core i7 gives a 36% higher VR performance
• 5G: Intel powers 5G end-to-end, only company with that capability, collaborating with partners and standards with yearly modem cadence and breadth of portfolio: in cloud, in network (Nokia, cicso, at&t,…), in access network (telcos,…), wiresless tech (mmWave,…), devices (1st global 5G modem sampling ’17 with Intel 14nm tech), LTE on 14nm in 2018

Next up: Diane Bryant on data center


Plus see:
https://twitter.com/witeken/status/829741619339354112

 

[liahos1]

data center products getting priority access to new nodes going forward according to diane.

 

[witeken]

Diane Bryant: Data Center
TL;DR:

*Long-term double digit DCG growth through 2021 (assuming 5% enterprise decline)

*Data center first on new node on 7nm and 3rd wave of 10nm (10nm++), instead of PC!!!

*Operating margins from high-40s% to low-40s%

*Cloud important + networking

*AI

*Omni-path, Si photonics, 3D XPoint, rack-scale, etc.

• Enterprise expectations: from mid-single digit growth to now expecting mid-decline
• Operating profit will grow smaller than revenue because of investments (“data center first”)
• DCG wants to be first on new node!!!!!! But has to get around die size (= yield) issue with innovation
• Growth through 2021: 6% CPU CAGR and 25% non-CPU CAGR
• Private cloud 25% CAGR growth, but still small (20%)
• 2/3 of cloud growth is TAM expansion
• Business-cloud buy high ASP products (eg custom products) vs consumer cloud, with dedicated sales force (20% headcount increase), ASPs of the top 8-107 is up >30% because of that sales
• Networking: moving from legacy ARM, MIPS fixed function to Intel Architecture, 5G will accelerate that for ~2020, now #1 in wireless access infrastructure, DCG+PSG (including FPGA) has <20% market share, with a $19B TAM. Network has nice ASP growth (almost +20% in ’16) because of moving up the product stack. Broadwell SoC (Xeon D) has caused ASP uptick (vs. the previous Atom SoCs).
• Artificial Intelligence: fastest growing workload, 7% of data center workload in 2016. 60% is classic machine learning, 40% is deep learning (image and language mainly). Intel has 97% MSS of classic machine learning. Deep learning (training + inference): 91% MSS, but deep learning is matrix math, which is good for GPUs, so Intel needs to react. Customers ask for solution to have homogeneous solution. Intel = scale. AI on IA.
• End-to-end products for AI: Skylake (machine learning), Knights Mill (classic machine learning), BDW+Arria 10 (programmable, low-latency inference), Lake crest (best deep learning training perf)
• DCG TAM: 46B in 2016 (has 36% MSS) to 65B in 2021.
• Omni-path: 42% of 100G HPC market share in first year, will be 1B market in 2021
• Photonics: 9% MSS, will be 1.8B TAM in 2021 (100G, moving to 400G): Intel advantage is laser integration in IC, only becomes bigger advantage over time
• 3D XPoint DIMM ramp in 2018, to replace the DRAM market (no MSS at all today), delivers 40% lower cost, assuming single digit attach rate, 8B SAM
• Rack-scale design architecture

 

[witeken]

Alright, we're half-time .

Most interesting announcement was the "data center first". Makes a ton of sense, honestly, given their corporate strategy, although it's very, very, very, very surprising announcement given the die size and resulting yield challenges.

After the break: IoT, memory and finance.

 

[raghu78]

Its not surprising that data center products will move first to new nodes from 7nm. In the past Intel used to take 18 months to move their server CPUs to the latest node. Ivy bridge EP launched in Sep 2013 (18 months after Ivybridge desktop launched in Apr 2012) . Similarly Broadwell EP launched in Q2 2016 roughly 18 months after Broadwell launched in tablets in Q4 2014. The problem for Intel going forward is AMD have a competitive CPU core and a steady roadmap. Intel cannot afford to have this 18 month lag in moving server CPUs to the latest node. Intel is likely to move their Xeons to 10nm in 2019. AMD should be moving to GF 7nm in 2019. AMD does not have such a long lag in moving their server CPUs to a new process node. So its almost mandatory for Intel to move to the new nodes asap for their server CPUs. Between 2011 and 2016 Intel could afford to move their server CPUs to new nodes much later than their desktop CPUs as AMD's Bulldozer was a disaster. The problem for Intel when moving server CPUs to new nodes is going to be yields as their server dies are massive at 500 sq mm. AMD has found a innovative method of going MCM with small dies (< 200 sq mm) and could even afford to go up to 250-300 sq mm per die going forward.

The important question for Intel is did they have a roadmap for the next 2-3 years so we get an idea of what their plans are. Some amount of granularity atleast at a half year level as to when their next tock will launch. If Intel cannot say that even now about their 10nm roadmap then its a big red flag. Their 10nm process is supposed to ramp in H2 2017. So they should have information about their 10nm roadmap atleast now.

 

[Dresdenboy]

People seem to be a bit unhappy. Does anyone of you have a clue, why they react this way? I didn't follow the meeting.

 

[witeken]

People seem to be a bit unhappy. Does anyone of you have a clue, why they react this way? I didn't follow the meeting.

People are stupid .

If I were an investor, and wanna be safe long-term, Intel really is a great investment, tbh. Certainly at $35. You can't really do wrong with that. Also has 3% div yield.

 

[ShintaiDK]

People seem to be a bit unhappy. Does anyone of you have a clue, why they react this way? I didn't follow the meeting.

Its related to the Atom issues for the C2000.

 

[ShintaiDK]

 

[jpiniero]

That's sort of been the problem I think with Data Center; non-cloud is starting to decline. And the cloud companies are not married to x86 like the typical enterprise has been. So I suppose in that sense that moving data center to be first to keep the Cloud companies happy sort of makes sense.

 

[ShintaiDK]

That's sort of been the problem I think with Data Center; non-cloud is starting to decline. And the cloud companies are not married to x86 like the typical enterprise has been. So I suppose in that sense that moving data center to be first to keep the Cloud companies happy sort of makes sense.

Cloud is as locked to x86 as much as the Enterprise. There isn't anything else that will replace x86. And no, ARM is a complete joke in the area.

 

[witeken]

Rob Crooke: Memory
TL;DR:

*Leading edge floating gate tech enables higher density for a given 3D stack height + smaller form factor

*Multiple gens of 3DXP on roadmap, expect to be 5% of rev in ‘17

• Memory explanation: trade-off in the hierarchy between close to CPU, latency and cost
• TAM: DRAM = 48B, NVM = 41B. 3D XP between those, for real-time processing.
• 3D NAND: Intel key advantage is choice of floating gate technology vs. replacement gate tech of competitors --> Intel can put peripheral circuitry underneath the NAND cells --> enables smallest memory cell, highest density (at the given 3D stack height), lower die size.
• Architected for rapid cost and yield learning and transistions. Will leapfrog from 32 to 64 layers in 5 quarters!!
• Smaller form factor at platform level. E.g. Putting a full rack in a 1U unit by putting all the SSDs next to each other (I’ll add cute picture if I can).
• 3DXP: can use 2 terminals (instead of 3 of transistor). Directly writeable, to use as memory instead of just storage. Multiple gens in development.
• 3DXP: ramp expectation: less than 5% revenue in 2017.

Next up: Doug Davis (IoT)!

 

[raghu78]

Any product roadmap updates from Intel for 2018-2020 or is that going to have to wait till IDC 2017 ?

 

[jpiniero]

Cloud is as locked to x86 as much as the Enterprise. There isn't anything else that will replace x86.

Intel obviously doesn't share your views if they are getting this aggressive on data center. They need to keep things moving to keep the Cloud companies happy.

 

[Arachnotronic]

Its not surprising that data center products will move first to new nodes from 7nm. In the past Intel used to take 18 months to move their server CPUs to the latest node. Ivy bridge EP launched in Sep 2013 (18 months after Ivybridge desktop launched in Apr 2012) . Similarly Broadwell EP launched in Q2 2016 roughly 18 months after Broadwell launched in tablets in Q4 2014. The problem for Intel going forward is AMD have a competitive CPU core and a steady roadmap. Intel cannot afford to have this 18 month lag in moving server CPUs to the latest node. Intel is likely to move their Xeons to 10nm in 2019. AMD should be moving to GF 7nm in 2019. AMD does not have such a long lag in moving their server CPUs to a new process node. So its almost mandatory for Intel to move to the new nodes asap for their server CPUs. Between 2011 and 2016 Intel could afford to move their server CPUs to new nodes much later than their desktop CPUs as AMD's Bulldozer was a disaster. The problem for Intel when moving server CPUs to new nodes is going to be yields as their server dies are massive at 500 sq mm. AMD has found a innovative method of going MCM with small dies (< 200 sq mm) and could even afford to go up to 250-300 sq mm per die going forward.

The important question for Intel is did they have a roadmap for the next 2-3 years so we get an idea of what their plans are. Some amount of granularity atleast at a half year level as to when their next tock will launch. If Intel cannot say that even now about their 10nm roadmap then its a big red flag. Their 10nm process is supposed to ramp in H2 2017. So they should have information about their 10nm roadmap atleast now.

What are you saying? In the last thread where we had a discussion, I told you to expect Intel to move DCG/HEDT to leadership on nodes, and you vehemently disagreed with me telling name that they need to do small dies first.

Intel said at the conference that DCG product node leadership will begin with 10nm++, not 7nm.

This fits the roadmap that mikk showed us on this forum. Ice Lake-EP will possibly be out on 10nm++ in 2019 and Sapphire Rapid on 7nm in 2020.

 

[raghu78]

Exactly. Intel realizes the threat that TSMC and the ARM ecosystem along with AMD/GF poses to its cash cow DCG. Intel's 10nm and TSMC 7nm / GF 7nm are for all intents and purposes going to be competing in the same time frame when it comes to server products. Intel realizes that it cannot drag its feet going forward as there are hungry competitors who will move to the bleeding edge nodes at the earliest possible time. Intel is just acknowledging that their past methods will not work going forward.

 

[ShintaiDK]

Intel obviously doesn't share your views if they are getting this aggressive on data center. They need to keep things moving to keep the Cloud companies happy.

Sure they do. Its about growth factors. Its a reshuffle like always. Desktop lost to mobile previously as well.

The ARM server dreams is nothing but a pipedream. Company after company tries, company after company abandons it.