Slashdot Mirror
Intel To Manufacture Rival ARM Chips In Mobile Push
An anonymous reader writes: Chip maker Intel has entered an unlikely partnership with British semiconductor firm ARM in an effort to boost opportunities for its foundry business. The licensing agreement, which was confirmed at the Intel Development Forum in San Francisco, means that from 2017 Intel's Custom Foundry will manufacture ARM chips -- used by smartphone giants such as Apple, Qualcomm and Samsung. On the announcement of its latest earnings report, Intel was clear to highlight a shift in focus, away from the traditional PC market, to emerging areas such as the Internet of Things and mobile -- a sector dominated by one-time arch rival ARM. It seems that Intel has now decided to surrender to the latter's prominence in the field.
Do it for the mobile! Talk about late to the game though. Even Microsoft got the clue before this. I'm hoping Intel chokes on desktops enough to let AMD get another foot in the door. Competition is sorely needed on the desktop/laptop front.
a dusty old thing from buying some smaller firm something like 8 years ago.
if this is supposed to be a new economy, how come they still want my old fashioned money?
Intel doesn't just make CPUs, they make whole systems. A PC with an Intel chipset has Intel NICs, Intel SATA controllers, Intel PCIe bus controllers, Intel USB controllers etc. They actually do a whole lot more, including cellular modems, flash memory controllers and all sorts stuff that gets integrated into smartphone SoCs.
So even if the CPU core is ARM instead of Intel Atom, the rest of the system will be Intel custom hardware. Getting that accepted and widely used is even more important than getting their CPU ISA used in mobile devices.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
This is amusing in part because Intel made ARM processors from 1997-2006 (branded StrongARM and later XScale), but decided that ARM processors were silly and sold their ARM processor business in 2006. In hindsight, that was the worst possible timing since the mobile market started to take off shortly thereafter (the first iPhone was 2007. Oh well. At least they're no longer wasting their time trying to cram x86 processors into phones.
Translation: There is really really truly competition in the semiconductor industry. It is not being consolidated and constricted by western government and forced to provide preferential supply to the west, especially their intelligence agency, to give them an edge over the natural resources of Russia and others in the global trade.
I wonder if there will be a successor to ARM. I'm not a CPU architect, but some of the main differences seem that ARM can mix cores (a multi-core processor need not have all cores of the same type) but overall requires more consecutive operations to perform a function (CISC vs RISC), but do so without the same overhead microcode translation.
In the future I'd imagine we'll be seeing more hybrids between the two, like APU's where essentially a die has an x64 CISC-style core for general processing and an ARM RISC-style core for things like graphics processing etc, but I'm also hoping some new and exciting will come around in the next half-decade or so (quantum computing, perhaps?).
Robotic assembly lines make it relatively easy and quick to switch on production of just about anything requiring mass quantities. Scaling up is easier with robotics.
Speaking as someone who has spent a fair bit of time working with robotic assembly lines, I think you are overestimating the plug-and-play nature of them severely. Mass production does not require robots nor is it particularly made easier by their presence. The advantages of robots are that they can work in hazardous environments, they can lower unit costs in some (not all) cases by reducing labor costs, they can produce repeatable products, and they sometimes can work faster. Downsides include: High up front tooling costs, less flexible than humans, require substantial programming time/expertise, too expensive for low-medium volume, maintenance and repair, and high setup costs.
As a general proposition scaling up is not any easier with robotics than with people and generally not any faster either. In some cases it can actually be more difficult. There are advantages to automation but ease of scaling is generally not one of them.
I just read an article how an Intel product (the don't deserve more advertisement here but it is a measurement of energy) . I should have skipped to the end of the article where there was the price. $369... wow, I could get 10 Pis, heck how about more than 70 Pi zeros, or more than 30 C.H.I.P.
Not sure who they are marketing this to.
What monopoly? ARM is still designed by a number of semiconductor designers - Qualcomm, Freescale, Apple (PA Semi), TI, et al, and still fabbed at TSMC, Samsung, et al. Intel here is just one more fab added to the latter
Intel needs to be viewed as several businesses, one of which is their discrete CPU business, another being their flash memory, yet another being the McAfee software acquisition, and yet another something called the foundry business.
Foundry refers to having a business where you are simply the manufacturer of chips for other companies for their specific purpose without selling into the end market. These other companies contract to Intel to be able to build anything from a network chip to a graphics chip to a microcontroller or virtually anything else (besides memory), either as a standard product off-the-shelf, or as an application-specific integrated circuit. In order for Intel to make that happen, they need to provide the know-how to these manufacturers of chips either directly or through providers of chip intellectual property. This includes logic libraries (standard cells, hence my name), memory cells and compilers for SRAMs, analog I/O cells, mixed-signal like ADCs and DACs, PLLs, non-volatile storage, design rule decks for the process rules, and a few other things that constitute the building blocks of any chip.
Other foundries such as TSMC, Global Foundries, etc. have the same model, though Intel's foundry manufactures more of their own CPU (and other) products than for other folks. Intel decided to farm out some of that capacity to third parties and make additional money on any spare capacity they might have, particularly with their leadership in logic processes over other rivals in the discrete CPU business. One of the key aforementioned building blocks is the IP offered by ARM for CPUs, GPUs and bus interconnect. This ARM IP needs to be validated to work in their silicon process, and this is the essence of the deal - Intel's foundry customers would not do business with Intel without basic blocks like the CPU since ARM is essentially the most important embedded CPU architecture in chip design currently.
The way the summary comes out makes it sound like Intel is manufacturing chips for its competitor, but it isn't necessarily so since the Intel microarchitecture is very highly vertically integrated as a business with their discrete CPU division whereas ARM itself is just a provider of IP with their microarchitecture. Yes, in theory Intel foundry customers could be making chips to compete in some segments of the Intel discrete CPU business, but that business is still largely dominated in the server and desktop markets by Intel and its associated software ecosystem. In the same way, ARM dominates the handheld device markets where Intel has had very little comparative presence.
I can guarantee that Mr. Krzanich and the Intel board would never allow their foundry business to cannibalize their current core discrete CPU business for a "competitor" if they felt it was detrimental to their overall financial and operating picture. This ARM deal is a piece of a larger plan of maximizing their ROI on their very very expensive chip fabs in a market where they have typically had a lead in logic process technology at least one node ahead of their competitors historically. That advantage can be very important in mobile due to the cost and power savings vertical transistor process nodes now offer along with superior manufacturing capabilities as the scale of their other businesses has long demonstrated.
The fact that Intel is offering to manufacture ARM cores for their custom foundry customers is not new. In fact, there are some Altera FPGAs with embedded ARM cores being manufactured by Intel already. The important thing about this deal is that ARM limited will now provide Hard IP for Intel's process technology.
To understand the importance of this, you have to understand a little more about silicon design and manufacturing than the average Slashdotter. Suppose you are some random fab-less chip designer that builds semi-customized ARM SoCs, a company like Rockchip or Mediatek for example. Generally the way you put together your new SoC is you buy a license for the ARM CPU design, then you buy a license for a GPU design from someone else, then you license a USB controller... so on and so forth, until you have all the building blocks necessary to make your new chip. Then you plug them all together, simulate, fab, validate, and ship.
Those blocks come in two different forms, Hard IP and Soft IP. Soft IP is basically a netlist... its a big text file that lists every transistor in the design and the interconnections between every transistor in the design. Usually soft IP vendors will give you the RTL, which is a more human readable language like Verilog which you compile in to a netlist. Hard IP on the other hard, is more like a vector graphics drawing or a stencil. Hard IP lists every transistor, its x/y coordinates on the silicon, and the exact shape and route of the copper wires. The problem with hard IP is every silicon manufacturer uses different shapes and sizes for their transistors and connecting wires (this is called the process design for the foundry), so a given hard IP design can only be built by the foundry it was designed for.
There is a program called a synthesizer that takes the netlist from the soft IP and generates the layout for the hard IP given a bunch of input parameters that describe the target foundry's process design, rather incredible really. The problem is not every design is "fully synthesizable" for example anything involving high speed I/O or analog (aka the "PHY" layers for modern busses: PCIe, USB, eMMC, Ethernet, SATA etc.) In any case, the pieces of the design that can not be synthesized need to be drawn by hand (aka human hands) using CAD software. For things like CPUs, usually there are some critical pieces that are drawn by hand, because a good human engineer can design a better, more efficient layout than the synthesizer can, at much greater expense of course. So depending on what percentage of your design is not synthesized, switching from one foundry to another can turn out to be a lot of work! This is the important thing here, ARM is providing ready to go hard IP for Intel foundry, just like they do with TSMC already, so the technical barrier for an ARM SoC designer to use Intel foundry is now lower... potentially comparable to TSMC.
Depending on the amount of engineers you have and how sophisticated they are, you might design some of those blocks yourself. Up to the point of companies like Apple and Qualcomm where even the ARM CPU design is a custom implementation and doesn't bear much resemblance to the reference design from ARM limited.
For Intel, using Intel foundry is a non-issue since they have an army of engineers that for the most part they design every IP block themselves anyway. For companies like Apple and Qualcomm that also have armies of engineers switching to Intel foundry is not a technical issue, its about business decisions for them. The big news is the smaller companies that don't have as many resources to do custom design now have Intel foundry as a viable option.
Most semiconductor companies buy a license that only lets them use designs from ARM Holdings plc. A few have "ARM Architectual Licenses" that allow them to do far more custom work on the CPU side of the SoC. Qualcomm, Apple, NVIDIA, and Cavium are the some who've done major work with their architectural licenses. Others like Samsung (Exynos) are mainly doing relatively minor modifications to the Cortex-Axx series from ARM.
“Common sense is not so common.” — Voltaire
Short term, the best bet for competition in the PC market is probably AMD's upcoming Zen architecture. Even if it cannot completely match the current Intel offerings, it will put pressure on Intel to sell their stuff at more affordable prices.
C - the footgun of programming languages
I think things went seriously wrong for AMD with the Bulldozer architecture in late 2011. Before that, they were already at a disadvantage vs. Intel's Sandy Bridge architecture (released earlier in the same year), but AMD's Phenom II still sold well due to cheaper prices.
Unfortunately, Bulldozer did not bring much of a performance improvement over the Phenom II, despite improving the manufacturing to a 32nm process. In some benchmarks, it even performed worse than the Phenom II. So stagnation at AMD while Intel kept making progress (albeit slowly).
The Piledriver architecture brought only a minor improvement.
The following Steamroller and Excavator architectures look better and get passable reviews in the notebook market, but are only available as APUs with a few exceptions. The FX series for the desktop is still on Piledriver, and frankly it cannot compete with the latest Intels.
This will hopefully change when the Zen architecture is released in late 2016/early 2017.
C - the footgun of programming languages
Right now, there are not many alternatives, especially in the x86 world. AMD is a US company too, I would not bet on them resisting a national Security Letter.
Other countries:
The Chinese have the Longsoon, a MIPS-derived homegrown architecture. The Russians have the Elbrus processor family.
For the Elbrus, I found a story on qz.com (http://qz.com/419923/this-22-pound-made-in-russia-laptop-is-actually-pretty-useful/) about a computer with the Elbrus-4C CPU inside. Expensive ($7500), only for sale to corporations and somewhat obsolete compared to the latest western PCs.
I guess for China and Russia it will be worthwhile to use those in security-critical environments, but for others not so much.
C - the footgun of programming languages
Shared data and a machine that can crush the phone sized thing (both literally and figuratively) is the much better option. They are two diametrically opposed form factors even without getting into the underlying guts of the things.
A Pirate and a Puritan look the same on a balance sheet.