When I saw the title, I assumed this was Intel's way to get into ARM without losing face. But... no. It's Atom. They just won't give up on x86. Sorry Intel, nobody wants it, execpt for the few thousand Win8 tablets getting churned out every month.
You won't have to wait for the Rockchip SoC to get a handle on Silvermont versus ARM on a TSMC 28nm process -- the Intel SoFIA chips already announced and slated for shipment later this year (dual core) and early next year (quad core) will tell the same story. That's why Intel could work this deal with Rockchip -- they already had ported Silvermont to the TSMC process.
If Intel can sell Rockchip SoCs without any "contra-revenue" that will be a big help to Intel's bottom line. Actually, just a little help, since there's not a lot of revenue involved in low-end tablet chips, regardless of what core is inside.
Well I would like to see an explanation that makes sense. First of all Intel needs to get high volume with these SoCs to make a profit at all. That's not a given since x86 on Android has so far not been very successful, and the low end already has already several highly efficient CPUs (eg. Cortex-A7 and A17) that will compete with Silvermont.
Using 28nm TSMC means most of the profit will now go to TSMC, not to Intel. While the higher volume will offset the lower profits per chip, Intel now has additional costs, both in terms of porting to TSMC 28nm as well as even lower utilization of their 22nm fabs. Still think this is a win-win situation?
You have to look at this from a long-term perspective -- this is a *war* and sometimes there will be short-term tactical sacrifices to achieve the long-term goal.
In this case, what Intel is doing with SoFIA, is essentially recognizing that for the low-end smartphone and tablet markets, it needs a fully integrated part in order to compete. Not coincidentally, this market matters because it is not only likely the *largest* in terms of unit volume, but it is also the most robust in terms of growth.
Anyway, so Intel has 2 fundamental problems in mobile:
(1) X86-64 is currently a second class citizen in the world of Android. While Intel has spent a lot of money trying to enable this ecosystem for X86 (and it has done a pretty decent job), developers ultimately don't care to optimize for architectures that are the minority. AMD had a very difficult time convincing developers to use its proprietary X86 extensions and AMD had reasonable market share. To remedy this, Intel needs lots of volume.
(2) Intel, unlike Qualcomm/MediaTek/etc. has traditionally not focused on bringing down the bill of materials cost of its platforms. This is why Bay Trail requires significant BoM-equalizing contra-revenue to be saleable. On top of that, Intel's products have typically lacked the integration that its competitors have, further damaging its competitiveness.
So, to solve (1), Intel needs to go broadly and deeply, and it needs to do so ASAP. Every moment Intel loses is a moment that an Android developer doesn't bother to optimize for X86. That is why Intel is offsetting the bill of materials issues with contra-revenue with Bay Trail, and why it is so keen to get these TSMC 28nm products out (they are designed to not require contra-revenue).
To solve (2), Intel needs an integrated modem. The Infineon Wireless team that Intel acquired was likely working on a whole pipeline of baseband/RF chips on TSMC's processes. So, in order to integrate a modem, Intel couldnt' spend additional time/money to port the WIP designs to its own 22nm process -- it needed to get something out now. So, they ported over the Atom core to 28nm and are building the first iteration of integration-competitive, BoM-competitive parts on TSMC's 28nm process since that's where the modem is already.
Now, in the short term (2015), this means that Intel hands over the foundry margin, but it no longer has to pay the enormous BoM offset (~$20/unit per Intel CFO), so net/net, the SoFIA chips should be marginally profitable rather than money-losers per unit. However, you correctly note that Intel doesn't benefit much here from a profit standpoint and TSMC wins.
But do note now that with market share for IA established, the partnerships with the OEMs stronger, and with time to port over all of the relevant IPs, Intel will bring these parts to 14nm during early 2016 and beyond. This means now that Intel collects the foundry margin, doesn't pay too much for IP royalties (using its own CPU core), and has a manufacturing lead over everybody else.
At this point in time, even if Intel outsourced the SOC integration to Rockchip for some low-end SKUs, it would still in effect collect foundry margin AND the margin of an IP vendor (ala ARM). This is extremely profitable, especially as the 14nm factories will have pumped out 10's of billions of dollars worth high end Broadwells and Skylakes to cover most if not all of the depreciation of the factories by then.
At this point, if Intel can deliver competitive products, it will have a cost structure edge over all of the fabless players, thus starting the chain reaction to ultimate long-term business success in this area.
It's a good strategy, but for it to work, Intel needs to execute, something its mobile group has had difficulty with in the past. Time will tell, however.
Intel wants access to ARM IP without paying for it. And access to design knowledge that offers a quick learning curve.
Just like Intel wanted and took Cyrix IP, wanted NexGen IP, wanted AMD IP and got away with the K6 feature set and design schedules, wanted DEC Alpha IP; took some you tell me, the close scrutiny of processor and platform innovations that come from a leading design enablement.
Intel so often the fast follower is certainly skilled in the subtle techniques of reuse.
Today ARM World presents a double edged sword.
Because if ARM we're to sell some kind of design license to Intel, at this point in the tragedy that is Intel confidence placements into ARM Inc., there would be a massive licensee lawsuit targeting ARM Ltd.
It's bad enough ARM over subscribes the design license that presses customers into purchasing the architectural license; although a necessity for competing in Intel time and Intel space.
Not to mention ARM Ltd. obsolescing designs in the pipe on rampant intellectual property introductions, as of late, that does wipe out cluster RISK development, enabling a contrived game of leap frog without fully considering the constituent effects. Or offering benefit of a coordinated plan to defray them?
Is ARM selfish or running scared? Acting like some other enterprise we know? Where (you fill in company name), a take it or leave it attitude so often one indicator there will be new entrants soon.
Finally, one must consider the regulatory question of two monopoly intellectual property holders having combined horizontally. Conversely Intel is aimed too displace some open industry volume, at TSMC, that Intel cannot afford to fabricate on their own in house marginal cost structure.
Times are interesting and the developing world is indeed becoming a very convoluted place.
Intel is unlikely to use ARM GPU IP, but Intel's own IP is out of the question for this generation as well given that Intel's major graphics overhauls happen at 14nm (and the current Gen 7 GPU from Intel is very poor).
My guess is that the entire SoFIA lineup, as with the 22nm Merrifield/Moorefield platforms, will use Imagination IP. Intel has experience writing drivers for ImgTec AND Intel is Imagination's largest shareholder.
Do we know who is actually doing the chip design. Is it really rockchip ?
If it is Intel, then Intel would need a Mali license.
Also, if by whatever way, Mali could theoritcally be an option for this chip, does ARM have any history whatsoever in Mali x86 graphics drivers ?
Also it seems a strange chip design alliance. Rockchip will have zero experience in designing a chip using Intel's internal chip bus structure, all of rockchips designs, and any IP they bring, will have been built around the standard ARM interconnect bus.
All of the IP will be driven in the past using ARM designed/customized and debugged drivers.
Adding to some of the observation and queries here:
Intel ARM crush began 2010 on roll out of three standard prototype platforms; Q7, Kontron MXM3, OCP Open Hub. These deployments enable system designers to evaluate the various processor options for considering all the standard criteria; product features, applications fit, power, performance, SDK, processor availability, supply and price. Intel played a leading role in the envisioning of these 'open' prototype platforms, each of which can also be used for low volume production.
In relation to ARM cluster trailing in process advantage, Intel knew at two lithography generations ahead, in conjunction with other criteria that determines a design win, Intel would win most of the time when compared with all the competitive processor option's in these 'open' prototype environments. Intel would win on platform evaluation and supply criteria Intel has always won on. At two process nodes ahead Intel would be chosen on price performance for no batteries required. Where batteries are required the question is what is good enough for power, ease of implementation and platform bill of materials aiming for some target market price? In both platform scenarios hardware development begins, and on the parallel software development Intel locks in that design win for the foreseeable future.
On optimizing any design Intel has always operated on the 80% rule. As long as the design is 80% there it can be re-architected and spun in next generation. Often times Intel designs are also ahead of their process economics. Intel on fabrication advantage will live with overly large dice for a generation then shrink. There is a rule when adding new feature blocks, cannot add greater than 10% area in generation. Yet this rule has not always been adhered too in desktop and server.
For tablet and phone, total power, package and footprint requirement of the system IDM does raise new hurdles for Intel, in terms of what their process can deliver, on a leading edge cost to price ratio. Which mean's marginal revenue in Intel's pocket, or lack of marginal revenue on Intel's marginal cost increase.
Intel's ARM trap had been set.
In same period, 2010, Intel would begin flooding supply channels with three consecutive generations of product; Nehalem, SandyBridge, IvyBridge, delivering a tragedy of the commons. An environment where Intel consumes system house manpower financial resource to work on anything but Intel platforms, and channel's financial capability to floor competitive processors for open market builds. Not envisioning nor understanding the realities of this Intel competitive environment, by 2012, ARM experienced their Dunkirk where cluster development addressing Android Chromebook and white book tablet market were, except for IDM strongholds, blown out of the water and gunned down attempting too establish beachheads. ARM discovers the competitive sale's stopper of the Intel package deal; commit to Intel volume purchase of premium processor's and get the popcorn free or at average price reduction.
Town common's is not necessarily overgrazed in prior example but is certainly over fertilized with Intel products, that cannot be effectively broken down on the industry's competitive organics. In other words, Intel can move product into the supply system, but the supply system is incapable of discharging that product, in Intel time, the competitive innovation's then tend too suffer on many forms of neglect.
ARM is also consumed by Intel relations. It is still hard to say whether Intel relations inside ARM where placed their clandestinely, or in that continental tradition of allowing some of the enemy in seeking accord; having to do with a long flawed strategy that ARM might just sell Intel something. ARM is duped and played into a number of Intel traps; all foretold on historical record of x86 war and competitive escalations of the 1990's and 2000's. Then there is the real pisser of some of these Intel relations, in ARM, subtly raising platform entry hurdles for server licensees.
From 2011 to now, Intel's base strategy has been the classic crush, to surround ARM cluster in their strongholds and squeeze on gravity of Intel processor's produced for supply over the processor offerings of ARM cluster licensees produced for demand. In this environment Intel surrounds and squeezes while taking small bites and looking for tapers into ARM product spaces. Where ARM fails too protect and to organize counter offensives, industrial embedded for example, Intel simply walks right in too reclaim lost territory. Engine system control seems to be next on Intel aim to recapture lost territory in automotive.
In robotics market ARM did not even stage a fight, ceding the robot's brains to Intel iCore and Atom happy with motor control. On that brainiac margin disparity Intel will eventually take the whole machine.
Replacements and substitutes; Intel open prototype platform's began supporting substitution, as a lure, yet most development opportunities in ARM world, and the open x86 market, support processor platform replacements. In this environment competitive advantage is defined by a processor's unique design features and interface into the system boards. Producers are held by processor interface and the promise of their future features validated on supply growth and sustainable road map. Qualcomm and Samsung are superb at this quasi open system game. Apple and Samsung are superb in their closed system games.
Intel licensing Atom core(s) to Rock Chip makes sense then for some very basic Intel reasons. First Intel loves to observe and borrow from others supporting internal developments. The classic anomaly is most recently observed in ARM server, where VIA in x86, followed by ARM licensee's fell for that old trick of entrusting their nascent micro server efforts to the research and development labs of Intel Dealers. On that ARM licensee investment in time, finance, manpower and effort, Intel learned every step of the way through their network of engineers who owe their long term employment preservation to Intel Network. This is always the case, even in this Rock Chip example, where HP procured Rock Chip ARM for low price Android Slate 7, that obviously woke up Intel. Intel will immediately gain access to ARM platform development and intellectually property reference, and will take that knowledge into account for deciding next chip architecture and platform design move.
That is one strategic reason for this deal. When design knowledge is one of the most sustainable competitive advantages.
Too sustain in these Intel Dealer developments, again citing HP as the lesser of the more notorious example, TI ARM + KeyStone DSP and AMD x86 + APU cartridge's offer an end use market applications performance advantage for HP MoonShot that is less easily traded on processor supply and price, as procurement negotiating chip, by Intel Network. Sustaining a design win becomes a competitive function of tools and performance for the application requirement validated by a supply on time, in volumes required, supporting an alternative market that demands that platform on its utility benefits which is the clincher.
Extent Intel can counter competitive procurement's on shenanigans of supply mishap, allocation melodrama including restraining access to premium parts and forward design data becomes much less when four competitive aspects can be met: tool, performance, applications and end market demand. Competitive platform developments among a few Intel Dealers are easily targeted and shot down by Intel. Where there is no competitive channel mass there is no competitive market leverage.
So where does the Intel Rock Chip deal go from here? First question too consider is whether the Intel Rock Chip SOC will be a classic x86 replacement. So similar to AMD x86 or VIA x86 offering some feature set that on desirability ties the system design producer too the design on its processor interface? Just like in the ARM World. For application some ball or bump mount array that delivers the processor's superior feature set to the system design engineer.
The more novel move where Intel is incredibly cagey, often nefarious, will punish those who use this strategy against Intel, is too deliver, an x86 core substitute that can interface with Rock Chip ARM peripheral blocks. That is the Coup de'état on my continuous audit; intersection of x86 and ARM, for which Intel truly drools. Technically to deliver a core processor that on its interface takes ARM compliment peripheral blocks on a substitutes leading taper.
Intel is squeezing using all their classic crush techniques, but to crush eventually means too implode and to implode mean too puncture and to taper in on those same strategies of their classic internal assaults. Which have always been too dismantle competitor's from within them.
On tablet producer contra revenue support, Intel appears to price Bay Trail T along that short run's price = marginal revenue = marginal cost curve. Meaning Bay Trail Mobile and Desktop subsidizes Bay Trail T.
On a good day Bay Trail full run; T, M, D variants at 10,000,000 units average weighed price on grade split exact is $99.32; average marginal revenue $85.21; average marginal cost $13.11. That is a design production cost of 13 cents mm^2. Intel's hard manufacturing cost is about $6.55.
On a bad day Bay Trail full run; T, M, D variants at 10,000,000 units average weighed price on grade split exact $98.59; average marginal revenue $69.13; average marginal cost $29.47. That is a design production cost of 29 cents mm^2. Intel's hard manufacturing cost is about $14.73.
Taking into account Bay Trail T short run alone at 3,780,799 units, meaning 62% of tablet is actually Core, average weighed price on grade split exact is $32.46; average marginal revenue $22.76; average marginal cost $9.70. That is design production cost of 9 cents mm^2. Intel's hard manufacturing cost is about $4.85. At $5.50 China price Intel is selling some Bay Trail T below their fully burdened cost to produce.
At marginal revenue $22.76 which is not the first 'T 'chip produced nor is it the last 'T' chip produced, Intel's contra cost subsidy is minimally a marginal revenue loss of $83,681,092 for every 3,589,001 units produced.