Computed·Blg

Via re/code

-----

Can a silicon chip act like a human brain? Researchers at IBM say they’ve built one that mimics the brain better than any that has come before it.

In a paper published in the journal Science today, IBM said it used conventional silicon manufacturing techniques to create what it calls a neurosynaptic processor that could rival a traditional supercomputer by handling highly complex computations while consuming no more power than that supplied by a typical hearing aid battery.

The chip is also one of the biggest ever built, boasting some 5.4 billion transistors, which is about a billion more than the number of transistors on an Intel Xeon chip.

To do this, researchers designed the chip with a mesh network of 4,096 neurosynaptic cores. Each core contains elements that handle computing, memory and communicating with other parts of the chip. Each core operates in parallel with the others.

Multiple chips can be connected together seamlessly, IBM says, and they could be used to create a neurosynaptic supercomputer. The company even went so far as to build one using 16 of the chips.

The new design could shake up the conventional approach to computing, which has been more or less unchanged since the 1940s and is known as the Von Neumann architecture. In English, a Von Neumann computer — you’re using one right now — stores the data for a program in memory.

This chip, which has been dubbed TrueNorth, relies on its network of neurons to detect and recognize patterns in much the same way the human brain does. If you’ve read your Ray Kurzweil, this is one way to understand how the brain works — recognizing patterns. Put simply, once your brain knows the patterns associated with different parts of letters, it can string them together in order to recognize words and sentences. If Kurzweil is correct, you’re doing this right now, using some 300 million pattern-recognizing circuits in your brain’s neocortex.

The chip would seem to represent a breakthrough in one of the long-term problems in computing: Computers are really good at doing math and reading words, but discerning and understanding meaning and context, or recognizing and classifying objects — things that are easy for humans — have been difficult for traditional computers. One way IBM tested the chip was to see if it could detect people, cars, trucks and buses in video footage and correctly recognize them. It worked.

In terms of complexity, the TrueNorth chip has a million neurons, which is about the same number as in the brain of a common honeybee. A typical human brain averages 100 billion. But given time, the technology could be used to build computers that can not only see and hear, but understand what is going on around them.

Currently, the chip is capable of 46 billion synaptic operations per second per watt, or SOPS. That’s a tricky apples-to-oranges comparison to a traditional supercomputer, where performance is measured in the number of floating point operations per second, or FLOPS. But the most energy-efficient supercomputer now running tops out at 4.5 billion FLOPS.

Down the road, the researchers say in their paper, they foresee TrueNorth-like chips being combined with traditional systems, each solving problems it is best suited to handle. But it also means that systems that in some ways will rival the capabilities of current supercomputers will fit into a machine the size of your smartphone, while consuming even less energy.

The project was funded with money from DARPA, the Department of Defense’s research organization. IBM collaborated with researchers at Cornell Tech and iniLabs.

Via PCWorld

-----

Qualcomm is getting high on 64-bit chips with its fastest ever Snapdragon processor, which will render 4K video, support LTE Advanced and could run the 64-bit Android OS.

The new Snapdragon 810 is the company’s “highest performing” mobile chip for smartphones and tablets, Qualcomm said in a statement. Mobile devices with the 64-bit chip will ship in the first half of next year, and be faster and more power-efficient. Snapdragon chips are used in handsets with Android and Windows Phone operating systems, which are not available in 64-bit form yet.

The Snapdragon 810 is loaded with the latest communication and graphics technologies from Qualcomm. The graphics processor can render 4K (3840 x 2160 pixel) video at 30 frames per second, and 1080p video at 120 frames per second. The chip also has an integrated modem that supports LTE and its successor, LTE-Advanced, which is emerging.

The 810 also is among the first mobile chips to support the latest low-power LPDDR4 memory, which will allow programs to run faster while consuming less power. This will be beneficial, especially for tablets, as 64-bit chips allow mobile devices to have more than 4GB of memory, which is the limit on current 32-bit chips.

The quad-core chip has a mix of high-power ARM Cortex-A57 CPU cores for demanding tasks and low-power A53 CPU cores for mundane tasks like taking calls, messaging and MP3 playback. The multiple cores ensure more power-efficient use of the chip, which helps extend battery life of mobile devices.

The company also introduced a Snapdragon 808 six-core 64-bit chip. The chips will be among the first made using the latest 20-nanometer manufacturing process, which is an advance from the 28-nm process used to make Snapdragon chips today.

Qualcomm now has to wait for Google to release a 64-bit version of Android for ARM-based mobile devices. Intel has already shown mobile devices running 64-bit Android with its Merrifield chip, but most mobile products today run on ARM processors. Qualcomm licenses Snapdragon processor architecture and designs from ARM.

Work for 64-bit Android is already underway, and applications like the Chrome browser are already being developed for the OS. Google has not officially commented on when 64-bit Android would be released, but industry observers believe it could be announced at the Google I/O conference in late June.

Qualcomm spokesman Jon Carvill declined to comment on support for 64-bit Android. But the chips are “further evidence of our commitment to deliver top-to-bottom mobile 64-bit leadership across product tiers for our customers,” Carvill said in an email.

Qualcomm’s chips are used in some of the world’s top smartphones, and will appear in Samsung’s Galaxy S5. A Qualcomm executive in October last year called Apple’s A7, the world’s first 64-bit mobile chip, a “marketing gimmick,” but the company has moved on and now has five 64-bit chips coming to medium-priced and premium smartphones and tablets. But no 64-bit Android smartphones are available yet, and Apple has a headstart and remains the only company selling a 64-bit smartphone with its iPhone 5S.

The 810 supports HDMI 1.4 for 4K video output, and the Adreno 430 graphics processor is 30 percent faster on graphics performance and 20 percent more power efficient than the older Adreno 420 GPU. The graphics processor will support 55-megapixel sensors, Qualcomm said. Other chip features include 802.11ac Wi-Fi with built-in technology for faster wireless data transfers, Bluetooth 4.1 and a processing core for location services.

The six-core Snapdragon 808 is a notch down on performance compared to the 810, and also has fewer features. The 808 supports LTE-Advanced, but can support displays with up to 2560 x 1600 pixels. It will support LPDDR3 memory. The chip has two Cortex-A57 CPUs and four Cortex-A53 cores.

The chips will ship out to device makers for testing in the second half of this year.

Via SlashGear

-----

Qualcomm is readying a new kind of artificial brain chip, dubbed neural processing units (NPUs), modeling human cognition and opening the door to phones, computers, and robots that could be taught in the same ways that children learn. The first NPUs are likely to go into production by 2014, CTO Matt Grob confirmed at the MIT Technology Review EmTech conference, with Qualcomm in talks with companies about using the specialist chips for artificial vision, more efficient and contextually-aware smartphones and tablets, and even potentially brain implants.

According to Grob, the advantage of NPUs over traditional chips like Qualcomm’s own Snapdragon range will be in how they can be programmed. Instead of explicitly instructing the chips in how processing should take place, developers would be able to teach the chips by example.

“This ‘neuromorphic’ hardware is biologically inspired – a completely different architecture – and can solve a very different class of problems that conventional architecture is not good at,” Grob explained of the NPUs. “It really uses physical structures derived from real neurons – parallel and distributed.”

As a result, “this is a kind of machine that can learn, and be programmed without software – be programmed the way you teach your kid” Grob predicted.

In fact, Qualcomm already has a learning machine in its labs that uses the same sort of biologically-inspired programming system that the NPUs will enable. A simple wheeled robot, it’s capable of rediscovering a goal location after being told just once that it’s reached the right point.

However it’s not only robots that can learn which will benefit from the NPUs, Qualcomm says. “We want to make it easier for researchers to make a part of the brain” Grob said, bringing abilities like classification and prediction to a new generation of electronics.

That might mean computers that are better able to filter large quantities of data to suit the particular needs of the user at any one time, smartphone assistants like Google Now with supercharged contextual intuition, and autonomous cars that can dynamically recognize and understand potential perils in the road ahead.

The first partnerships actually implementing NPUs in that way are likely to come in 2014, Grob confirmed, with Qualcomm envisaging hugely parallel arrays of the chips being put into practice to model how humans might handle complex problems.

Via ars technica

-----

A desktop PC used to need a lot of different chips to make it work. You had the big parts: the CPU that executed most of your code and the GPU that rendered your pretty 3D graphics. But there were a lot of smaller bits too: a chip called the northbridge handled all communication between the CPU, GPU, and RAM, while the southbridge handled communication between the northbridge and other interfaces like USB or SATA. Separate controller chips for things like USB ports, Ethernet ports, and audio were also often required if this functionality wasn't already integrated into the southbridge itself.

As chip manufacturing processes have improved, it's now possible to cram more and more of these previously separate components into a single chip. This not only reduces system complexity, cost, and power consumption, but it also saves space, making it possible to fit a high-end computer from yesteryear into a smartphone that can fit in your pocket. It's these technological advancements that have given rise to the system-on-a-chip (SoC), one monolithic chip that's home to all of the major components that make these devices tick.

The fact that every one of these chips includes what is essentially an entire computer can make keeping track of an individual chip's features and performance quite time-consuming. To help you keep things straight, we've assembled this handy guide that will walk you through the basics of how an SoC is put together. It will also serve as a guide to most of the current (and future, where applicable) chips available from the big players making SoCs today: Apple, Qualcomm, Samsung, Nvidia, Texas Instruments, Intel, and AMD. There's simply too much to talk about to fit everything into one article of reasonable length, but if you've been wondering what makes a Snapdragon different from a Tegra, here's a start.

Putting a chip together

There's no discussion of smartphone and tablet chips that can happen without a discussion of ARM Holdings, a British company with a long history of involvement in embedded systems. ARM's processors (and the instruction set that they use, also called ARM) are designed to consume very small amounts of power, much less than the Intel or AMD CPUs you might find at the heart of a standard computer. This is one of the reasons why you see ARM chips at the heart of so many phones and tablets today. To better understand how ARM operates (and to explain why so many companies use ARM's CPU designs and instruction sets), we first must talk a bit about Intel.

Intel handles just about everything about its desktop and laptop CPUs in-house: Intel owns the x86 instruction set its processors use, Intel designs its own CPUs and the vast majority of its own GPUs, Intel manufactures its own chips in its own semiconductor fabrication plants (fabs), and Intel handles the sale of its CPUs to both hardware manufacturers and end users. Intel can do all of this because of its sheer size, but it's one of the only companies able to work this way. Even in AMD's heyday, the company was still licensing the x86 instruction set from Intel. More recently, AMD sold off its own fabs—the company now directly handles only the design and sale of its processors, rather than handling everything from start to finish.

ARM's operation is more democratized by design. Rather than making and selling any of its own chips, ARM creates and licenses its own processor designs for other companies to use in their chips—this is where we get things like the Cortex-A9 and the Cortex-A15 that sometimes pop up in Ars phone and tablet reviews. Nvidia's Tegra 3 and 4, Samsung's Exynos 4 and 5, and Apple's A5 processors are all examples of SoCs that use ARM's CPU cores. ARM also licenses its instruction set for third parties to use in their own custom CPU designs. This allows companies to put together CPUs that will run the same code as ARM's Cortex designs but have different performance and power consumption characteristics. Both Apple and Qualcomm (with their A6 and Snapdragon S4 chips, respectively) have made their own custom designs that exceed Cortex-A9's performance but generally use less power than Cortex-A15.

The situation is similar on the graphics side. ARM offers its own "Mali" series GPUs that can be licensed the same way its CPU cores are licensed, or companies can make their own GPUs (Nvidia and Qualcomm both take the latter route). There are also some companies that specialize in creating graphics architectures. Imagination Technologies is probably the biggest player in this space, and it licenses its mobile GPU architectures to the likes of Intel, Apple, and Samsung, among others.

Chip designers take these CPU and GPU bits and marry them to other necessary components—a memory interface is necessary, and specialized blocks for things like encoding and decoding video and processing images from a camera are also frequent additions. The result is a single, monolithic chip called a "system on a chip" (SoC) because of its more-or-less self-contained nature.

There are two things that sometimes don't get integrated into the SoC itself. The first is RAM, which is sometimes a separate chip but is often stacked on top of the main SoC to save space (a method called "package-on-package" or PoP for short). A separate chip is also sometimes used to handle wireless connectivity. However, in smartphones especially, the cellular modem is also incorporated into the SoC itself.

While these different ARM SoCs all run the same basic code, there's a lot of variety between chips from different manufacturers. To make things a bit easier to digest, we'll go through all of the major ARM licensees and discuss their respective chip designs, those chips' performance levels, and products that each chip has shown up in. We'll also talk a bit about each chipmaker's plans for the future, to the extent that we know about them, and about the non-ARM SoCs that are slowly making their way into shipping products. Note that this is not intended to be a comprehensive look at all ARM licensees, but rather a thorough primer on the major players in today's and tomorrow's phones and tablets.

Apple

We'll tackle Apple's chips first, since they show up in a pretty small number of products and are exclusively used in Apple's products. We'll start with the oldest models first and work our way up.

The Apple A4 is the oldest chip still used by current Apple products, namely the fourth generation iPod touch and the free-with-contract iPhone 4. This chip marries a single Cortex A8 CPU core to a single-core PowerVR SGX 535 GPU and either 256MB or 512MB of RAM (for the iPod and iPhone, respectively). This chip was originally introduced in early 2010 with the original iPad, so it's quite long in the tooth by SoC standards. Our review of the fifth generation iPod touch shows just how slow this thing is by modern standards, though Apple's tight control of iOS means that it can be optimized to run reasonably well even on old hardware (the current version of iOS runs pretty well on the nearly four-year-old iPhone 3GS).

Next up is the Apple A5, which despite being introduced two years ago is still used in the largest number of Apple products. The still-on-sale iPad 2, the iPhone 4S, the fifth-generation iPod touch, and the iPad mini all have the A5 at their heart. This chip combines a dual-core Cortex A9 CPU, a dual-core PowerVR SGX 543MP2 GPU, and 512MB of RAM. Along with the aforementioned heavy optimization of iOS, this combination has made for quite a longevous SoC. The A5 also has the greatest number of variants of any Apple chip: the A5X used the same CPU but included the larger GPU, 1GB of RAM, and wider memory interface necessary to power the third generation iPad's then-new Retina display, and a new variant with a single-core CPU was recently spotted in the Apple TV.

Finally, the most recent chip: the Apple A6. This chip, which to date has appeared only in the iPhone 5, marries two of Apple's custom-designed "Swift" CPU cores to a triple-core Imagination Technologies PowerVR SGX 543MP3 GPU and 1GB of RAM, roughly doubling the performance of the A5 in every respect. The CPU doubles the A5's performance both by increasing the clock speed and the number of instructions-per-clock the chip can perform relative to Cortex A9. The GPU gets there by adding another core and increasing clock speeds. As with the A5, the A6 has a special A6X variant used in the full-sized iPad that uses the same dual-core CPU but ups the ante in the graphics department with a quad-core PowerVR SGX 554MP4 and a wider memory interface.

Apple SoCs all prioritize graphics performance over everything else, both to support the large number of games available for the platform and to further Apple's push toward high-resolution display panels. The chips tend to have less CPU horsepower and RAM than the chips used in most high-end Android phones (Apple has yet to ship a quad-core CPU, opting instead to push dual-core chips), but tight control over iOS makes this a non-issue. Apple has a relative handful of iOS devices it needs to support, so it's trivial for Apple and third-party developers to make whatever tweaks and optimizations they need to keep the operating system and its apps running smoothly even if the hardware is a little older. Whatever you think of Apple's policies and its "walled garden" approach to applications, this is where the tight integration between the company's hardware and software pays off.

Knowing what we do about Apple's priorities, we can make some pretty good educated guesses about what we'll see in a hypothetical A7 chip even if the company never gives details about its chips before they're introduced (or even after, since we often have to rely on outfits like Chipworks to take new devices apart before we can say for sure what's in them).

On the CPU side, we'd bet that Apple will focus on squeezing more performance out of Swift, whether by improving the architecture's efficiency or increasing the clock speed. A quad-core version is theoretically possible, but to date Apple has focused on fewer fast CPU cores rather than more, slower ones, most likely out of concern about power consumption and the total die size of the SoC (the larger the chip, the more it costs to produce, and Apple loves its profit margins). As for the GPU, Imagination's next-generation PowerVR SGX 6 series GPUs are right around the corner. Since Apple has used Imagination exclusively in its custom chips up until now, it's not likely to rock this boat.

Qualcomm

Qualcomm is hands-down the biggest player in the mobile chipmaking game right now. Even Samsung, a company that makes and ships its own SoCs in the international versions of its phones, often goes with Qualcomm chips in the US. With this popularity comes complexity: Wikipedia lists 19 distinct model numbers in the Snapdragon S4 lineup alone, and those aren't even Qualcomm's newest chips. So we'll pick four of the most prominent to focus on, since these are the ones you're most likely to see in a device you could buy in the next year or so.

Let's start with the basics: Qualcomm is the only company on our list that creates both its own CPU and GPU architectures, rather than licensing one or the other design from ARM or another company. Its current CPU architecture, called "Krait," is faster clock-for-clock than ARM's Cortex A9 but slower than Cortex A15 (the upside is that it's also more power-efficient than A15). Its GPU products are called "Adreno," and they actually have their roots in a mobile graphics division that Qualcomm bought from AMD back in 2009 for a scant $65 million. Both CPU and GPU tend to be among the faster products on the market today, which is one of the reasons why they're so popular.

The real secret to Qualcomm's success, though, is its prowess in cellular modems. For quite a while, Qualcomm was the only company offering chips with an LTE modem integrated into the SoC itself. Plenty of phones make room for separate modems and SoCs, but integrating the modem into the SoC creates space on the phone's logic board, saves a little bit of power, and keeps OEMs from having to buy yet another chip. Even companies that make their own chips use Qualcomm modems—as we noted, almost all of Samsung's US products come with a Qualcomm chip, and phones like the BlackBerry Z10 use a Qualcomm chip in the US even though they use a Texas Instruments chip abroad. Even Apple's current iPhones use one or another (separate) Qualcomm chips to provide connectivity.

Add these modems to Qualcomm's competitive CPUs and GPUs, and it's no wonder why the Snapdragon has been such a success for the company. Qualcomm will finally start to see some real challenge on this front soon: Broadcom, Nvidia, and Intel are all catching up and should be shipping their own LTE modems this year, but for now Qualcomm's solutions are established and mature. Expect Qualcomm to continue to provide connectivity for most devices.

Let's get to the Snapdragon chips themselves, starting with the oldest and working our way up. Snapdragon's S4 Plus, particularly the highest-end model (part number MSM8960), combines two Krait cores running at 1.7GHz with an Adreno 225 GPU. This GPU is roughly comparable to the Imagination Technologies GPU in Apple's A5, while the Krait CPU is somewhere between the A5 and the A6. This chip is practically everywhere: it powers high-end Android phones from a year or so ago (the US version of Samsung's Galaxy S III) as well as high-end phones from other ecosystems (Nokia's Lumia 920 among many other Windows phones, plus BlackBerry's Z10). It's still a pretty popular choice for those who want to make a phone but don't want to spend the money (or provide the larger battery) for Qualcomm's heavy-duty quad-core SoCs. Look for the S4 Plus series to be replaced in mid-range phones by the Snapdragon 400 series chips, which combine the same dual-core Krait CPU with a slightly more powerful Adreno 305 GPU (the HTC First is the first new midrange phone to use it. Others will likely follow).

Next up is the Snapdragon S4 Pro (in particular, part number APQ8064). This chip combines a quad-core Krait CPU with a significantly beefed up Adreno 320 GPU. Both CPU and GPU trade blows with Apple's A6 in our standard benchmarks, but the CPU is usually faster as long as all four of its cores are actually being used by your apps. This chip is common in high-end phones released toward the end of last year, including such noteworthy models as LG's Optimus G, the Nexus 4, and HTC's Droid DNA. It's powerful, but it can get a little toasty: if you've been running the SoC full-tilt for a while, the Optimus G's screen brightness will automatically turn down to reduce the heat, and the Nexus 4 will throttle the chip and slow down if it's getting too hot.

The fastest, newest Qualcomm chip that's actually showing up in phones now is the Snapdragon 600, a chip Qualcomm unveiled at CES back in January. Like the S4 Pro, this Snapdragon features a quad-core Krait CPU and Adreno 320 GPU, but that doesn't mean they're the same chip. The Krait in the Snapdragon 600 is a revision called "Krait 300" that both runs at a higher clock speed than the S4 Pro's Krait (1.9GHz compared to 1.7GHz) and includes a number of architectural tweaks that make it faster than the original Krait at the same clock speed. The Snapdragon 600 will be coming to us in high-end phones like the US version of Samsung's Galaxy S4, HTC's One, and LG's Optimus G Pro. Our benchmarks for the latter phone show the Snapdragon 600 outdoing the S4 Pro by 25 to 30 percent in many tests, which is a sizable step up (though the Adreno 320 GPU is the same in both chips).

Finally, look ahead to the future and you'll see the Snapdragon 800, Qualcomm's next flagship chip that's due in the second quarter of this year. This chip's quad-core Krait 400 CPU again introduces a few mild tweaks that should make it faster clock-for-clock than the Krait 300, and it also runs at a speedier 2.3GHz. The chip sports an upgraded Adreno 330 GPU that supports a massive 3840×2160 resolution as well as a 64-bit memory interface (everything we've discussed up until now has used a 32-bit interface). All of this extra hardware suggests that this chip is destined for tablets rather than smartphones (a market segment where Qualcomm is less prevalent), but this doesn't necessarily preclude its use in high-end smartphones. We'll know more once the first round of Snapdragon 800-equipped devices are announced.

Qualcomm is in a good position. Its chips are widely used, and its roadmap evolves at a brisk and predictable pace. Things may look less rosy for the company when competing LTE modems start to become more common, but for now it's safe to say that most of the US' high-end phones are going to keep using Qualcomm chips.

Samsung

Samsung has three-or-so chips that are currently shipping in its phones and tablets. The first (and oldest) of the three is the Exynos 4 Quad, which powers the Galaxy Note 10.1, Galaxy Note 8.0, Galaxy Note II, and international versions of the Galaxy S III. This particular variant includes four Cortex A9 CPU cores and an ARM Mali-400 GPU. Neither is cutting edge, but the GPU performance is better than Nvidia's Tegra 3 and the CPU performance is fairly similar (given similar clock speeds, anyway).

The other chips are both from the Exynos 5 series, but they're both quite different from each other. The first is the relatively straightforward Exynos 5 Dual, which powers both the Nexus 10 tablet and Samsung's $249 ARM Chromebook. This chip combines two ARM Cortex A15 cores with ARM's Mail-T604 GPU, and the result is the fastest GPU performance in any Android tablet at the moment and the fastest CPU performance in any ARM-based device, period. (This will quickly stop being the case as other A15-based devices start hitting the market this year). The chip is a bit more power-hungry than its Cortex A9-based predecessor and other designs from Apple and Qualcomm, but manufacturing process advancements absorb most of this penalty and Exynos 5 Dual devices still end up with decent battery life overall.

Finally, we have the Exynos 5 Octa, which is coming to market first in the international version of the forthcoming Galaxy S 4. This SoC is generally said to have eight CPU cores, and while this is not technically untrue, we've already pointed out that not all of these cores are created equal. The SoC combines four Cortex A15 cores for performance and four Cortex A7 cores that can run all of the same code, but much more slowly. Tasks that don't need a ton of CPU power can execute on the A7 cores, and tasks that do can execute on the A15s, but it's unlikely that all eight cores can be active at the same time. This chip's maximum CPU performance, then, will be more in line with a quad-core Cortex A15 chip like Nvidia's Tegra 4.

The Octa also ditches ARM's GPU designs for one by Imagination Technologies, a triple-core PowerVR SGX 544MP3. This is nearly identical to the 543MP3 used in Apple's A6, and the performance should be very similar. The only difference is that the 544MP3 supports Direct3D, a necessity if the Octa is to make its way into Windows phones or Windows RT tablets. Apple's competitors in the chip space are finally beginning to catch up with their GPU performance, something we couldn't have said of many chips even a year ago.

Samsung's chips have been known to appear in products from other companies, but they ship most frequently in Samsung's own phones, tablets, and (more recently) laptops. Samsung has the advantage of being a more integrated company than many of its competitors—not only does it make and sell its own phones and tablets, it also manufactures many of the components that appear in those devices, including the screens and the chips themselves. Nvidia and Qualcomm both typically outsource their chip production to TSMC, a company that also handles GPU production for AMD and Nvidia. Meanwhile, Apple (Samsung's biggest competitor in the mobile market) relies on Samsung for the production of the A5 and A6 chips that power its iOS devices.

Texas Instruments

Texas Instruments is an odd duck in this discussion. On the one hand, it provides chips for many prominent devices past and present, including Amazon's entire Kindle Fire, Samsung's Galaxy S II (and several other pre-Galaxy S III Samsung devices), and the international version of the BlackBerry Z10. On the other hand, TI has announced that it is exiting the market for smartphone and tablet SoCs and will be focusing on less-competitive, higher-margin markets—think embedded systems and factories. That doesn't mean it will be leaving the consumer market all of a sudden, just that it won't be devoting resources to new chips, and its existing chips will become more and more rare as time goes on.

The most common TI chips you'll find in products today belong to the OMAP4 series, which consists of three chips: the OMAP4430, the OMAP4460, and the OMAP4470. All use a dual-core Cortex A9 CPU (the higher the model number is, the higher the clock speed) alongside a single-core Imagination Technologies PowerVR SGX540 (in the 4430 and 4460) and a single-core PowerVR SGX544 (in the 4470). Two low-power ARM Cortex M3 cores are also included to help process background tasks while eating less battery.

The OMAP4's CPU performance is lower than the newer chips from Qualcomm or Nvidia, but like Apple's A5 it's generally good enough, especially when paired with Jelly Bean (or something like BlackBerry 10, which is optimized for it). The GPU performance, however, often lags behind not just newer chips, but also contemporaneous chips like the A5 or Nvidia's Tegra 3 (especially in the lower-end chips).

TI has one more consumer-targeted design in its pipeline, and it will probably be its last: the OMAP5. It uses the same basic setup as OMAP4, but everything has been upgraded: the two Cortex A9s have been exchanged for A15s, the Cortex M3s have been exchanged for M4s, and the GPU has been bumped to a dual-core PowerVR SGX544MP2 rather than the single-core version (the GPU's clock speed has also been increased to 532MHz, a little less than twice as fast as the PowerVR SGX544 in the OMAP4470). This should all add up to a GPU that's between three and four times as fast as its predecessor, always a welcome improvement.

OMAP5 is reportedly due in the second quarter of this year—so any day now. Even so, we haven't heard much about devices that will be using it. This silence may be because the product isn't actually on the market yet, but it may be the case that TI's anticipated withdrawal from the market has killed any chance this chip had to succeed. TI will probably be willing to cut buyers some pretty good deals, but if I had the option to buy a chip from a company with a well-charted roadmap (like Qualcomm or Nvidia) and a company that has announced its intent to completely abandon the consumer market, I know which one I'd choose.

Nvidia

The Tegra 3 is Nvidia's current SoC, and though it's getting a bit long in the tooth, it's still showing up in some relatively high-profile products. The chip uses four ARM Cortex A9 CPU cores and a custom-designed GPU made by Nvidia, which makes sense given its history as a graphics company. The SoC also includes a fifth low-power CPU core called a "companion core" designed to perform background tasks when your phone or tablet is idle, allowing the main CPU cores to power down and save your battery. There are a few different Tegra 3 variants, and they differ mostly in clock speed and memory bandwidth rather than core count.

The Tegra 3's CPU performs reasonably well, though at this point a quad-core Cortex A9 is going to feel slower than a dual-core CPU based on a newer architecture like the Cortex A15 simply because there aren't that many heavily threaded apps on phones and tablets these days. The GPU has also been surpassed by other offerings from Qualcomm, Apple, and Samsung, though the games actually available for Android today can usually be played without issue.

The Tegra 3 isn't as prevalent in phones and tablets as Qualcomm's chips, but it still powers plenty of Android and Windows RT devices. The Nexus 7, HTC One X+, Microsoft Surface, Asus VivoTab RT, and Asus Transformer Prime are all prominent devices using Nvidia silicon. The Ouya game console also uses a Tegra 3.

Tegra 3's successor is (unsurprisingly) called the Tegra 4, and the first devices to use it will be coming out in the next few months. Nvidia's own Project Shield gaming console will be one of the earliest to use it, but Vizio and Toshiba have both announced tablets that will use the chip as well. Tegra 4 uses the same basic configuration of CPU cores as Tegra 3—four cores, plus a low-power "companion core"—but trades the Cortex A9s for much more powerful Cortex A15s. The GPU is also much-improved and should go toe-to-toe with the GPU in Apple's iPad 4.

Tegra 4 is aimed at tablets and the very highest-end smartphones, but Nvidia is going a different route for mainstream smartphones. The Tegra 4i, due toward the end of this year, has the same basic GPU architecture as Tegra 4, but it uses a narrower memory interface (32-bit as opposed to 64-bit) and fewer cores (60 instead of 72). The CPU is also a little weaker—like Tegra 3, it's comes with four Cortex A9 CPU cores and one "companion core," but it's based on a revision of Cortex A9 called "Cortex A9 R4." The R4 promises higher performance than Cortex A9 at the same clock speed. Maximum clock speeds have also been increased significantly over Tegra 3, from 1.7GHz to 2.3GHz.

What will help Tegra 4i the most is the presence of an integrated LTE modem, the Icera i500. We've already talked about the benefits of having a modem integrated directly into the SoC itself, but this one has some unique aspects. The i500 is a "soft modem," which means that instead of having bits and pieces dedicated to communicating over specific bands or with specific standards, it has some general-purpose hardware that can be programmed to communicate over any of them as long as the rest of the hardware supports it. In theory, this would remove the need to build different models of a phone to serve different markets or different carriers. Both Tegra 4 and Tegra 4i also include a new imaging technology called "Chimera" that allows for always-on, real-time HDR photographs without the lag and blurriness that affects current HDR implementations.

Neither Tegra 4 variant is here yet, but that hasn't stopped Nvidia from talking about its plans for the more distant future. "Logan," a successor to Tegra 4 due in 2014, will use the same "Kepler" GPU architecture as Nvidia's current GeForce GPUs. Aside from the accompanying performance increases, this opens the door to GPU-assisted computing, which can be quite useful in workstation and server applications. Finally, 2015's "Parker" will incorporate Nvidia's first custom-designed ARM CPU, marking a move away from ARM's stock designs.

Nvidia's biggest challenge with all of these chips is going to be breaking into a market that others have largely cornered. Tegra 3 has made some inroads for them, but the biggest smartphone and tablet manufacturers (Apple and Samsung) already make their own chips, and (in the US at least) Qualcomm tends to be the go-to choice for most others. Still, with Texas Instruments leaving the market, we may soon see prominent companies that use its OMAP chips (Amazon, among many others) looking for an alternative. Nvidia can capitalize on this opening, especially if it can undercut Qualcomm on price (and according to Nvidia representatives I've spoken with, this is indeed the case).

Intel and AMD: x86 struggles to make the jump

We've talked almost exclusively about ARM-based products so far, but Intel, the 500-pound gorilla of the PC market, is still fighting to establish a reputation for making good tablet chips. Intel's current-generation products, the Ivy Bridge CPU architecture on the high end and the Clover Trail Atom platform on the low end, can't quite hit that necessary sweet spot between performance and power efficiency. Ivy Bridge tablets like Acer's Iconia W700 are still a little hot, a little heavy, a little expensive, and get only OK battery life. Clover Trail devices like Lenovo's ThinkPad Tablet 2 address all of these concerns, but their CPU and GPU performance is relatively low (GPU performance is especially bad) and the platform doesn't support Android.

Intel gets more interesting this year. Its Haswell chips should enable thinner, lighter tablets with better battery life than the Ivy Bridge models, while both the Clover Trail+ and Bay Trail Atom platforms look to deliver substantial gains in both CPU and GPU performance (Intel's cellular modems are also steadily improving, which helps). Intel's long-established relationships with the PC OEMs will ensure that both of these chips' architectures find their way into plenty of tablets, but we're still waiting for an Intel-powered smartphone to make its way to the US—so far, most Intel phones have been targeted toward "emerging markets."

AMD has also made a few moves in this direction: it has adapted its Bobcat netbook architecture into something called Hondo, which combines a dual-core CPU with an integrated Radeon GPU. By all reports, the CPU is in the same ballpark as Clover Trail's (the architecture is faster clock-for-clock, but Hondo runs at a lower clock speed than Clover Trail), while the GPU is a substantial step up. One of our main issues with Clover Trail tablets is that their GPUs deliver sometimes choppy UI and gaming performance, so improvements on this front are more than welcome.

No matter what the chip's virtues, though, its main problem is that most OEMs just aren't picking up what AMD is putting down. At our first Hondo briefing back in October of 2012, AMD played coy when asked about which devices Hondo would appear in. Since then, only two have been announced: one Windows 8 tablet apiece from Fujitsu and TV-turned-PC-maker Vizio. Bigger names are conspicuous in their absence, and unless AMD can develop a more convincing roadmap and get more people on board, it seems unlikely that its chips will amount to much.

AMD's first ARM processors are also coming in 2014, but they're targeted toward servers and not the consumer market. This (plus a number of recent hires) suggests that AMD could be looking to get into the ARM SoC game (and it could certainly handle the GPU despite selling its last mobile GPU division to Qualcomm, a move that seems short-sighted in retrospect). For now, its efforts remain focused squarely on the server room.

All of these chips have one potential trump card over the ARM chips we've talked about: x86. How important this architecture is to you will depend entirely on what you do: if you're a heavy user of Windows 8 or Windows desktop applications, x86 is a must-have because the ARM-based Windows RT can't run any of that stuff. If you prefer your tablets to be Android-flavored, Intel in particular has done a lot of work with Google to optimize Android for x86, and every Intel-powered Android phone or tablet we've seen has indeed performed pretty smoothly. Intel has also created something called "binary translation" to run most apps from the Google Play store without requiring much (if any) extra work on the part of the developers. Still, Android doesn't need x86 like Windows does, and if you're trying to build something on the cheap, Intel probably isn't your best option.

On Intel's end, the theory is that its manufacturing expertise will eventually outstrip its competitors' by so much that it will enable it to cram more performance into a smaller, more power-efficient chip. This is one possible outcome, though I think that companies like Apple and Samsung are going to be slow to move away from using their own chips in most of their mobile devices. If they can keep with performance that's "good enough," sticking with their own products might still be preferable to paying Intel for tablet and phone chips as they have for desktop and laptop chips for so long.

Where the market is going

There are other chipmakers in the world, but this has been a reasonably comprehensive look at the current offerings that you're most likely to see in most mid-to-high-end smartphones or tablets within the next year or so. Now that we've covered the products and their performance relative to each other, let's look at the market itself and the direction things seem to be going.

First, despite the number of players, the market for third-party chips is deceptively small. Look at Apple and Samsung, by far the most successful smartphone and tablet companies—Samsung often uses Qualcomm chips in its US phones, but otherwise both companies build and ship their own chips in their own products. Especially in Apple's case, this keeps a large, lucrative chunk of the market out of reach for companies that make only chips. Qualcomm, Nvidia, and the others have to fight it out for the rest.

As we've already discussed, Qualcomm is by far the largest third-party chipmaker in this game, and it has arrived at that position by delivering chips with good performance and versatile modems. It's the go-to choice for most Android and Windows Phone handset makers—currently, its quad-core chips are popular in the highest-end phones, while midrange phones like the HTC First can go with the slightly older, cheaper, but still dependable dual-core models. If you want to get your chips in your phones, Qualcomm is who you're fighting, if only because it's the biggest company you can fight.

That's exactly what Nvidia is trying to do with the Tegra 4i and its integrated Icera i500 modem: present a cheaper, all-in-one competitor to Qualcomm's mid-range and high-end products. Nvidia's biggest issue is actually similar to AMD's—it may be having some trouble convincing OEMs to use its new products. With Tegra 2 and Tegra 3, there's an impression that the company over-promised and under-delivered on things like performance and power consumption. Though it's early days yet for Tegra 4, we're still looking at a pretty short list of products that are confirmed to be using it, and they're all from pretty minor players. Everything I've seen so far about Tegra 4 (though admittedly seen through PR's rose-colored glasses) has been good, and TI's withdrawal from the market could be Nvidia's chance to snap up some new business.

Ultimately, TI's withdrawal shows how rough this market can be for any company that isn't Qualcomm. If the company that provides chips for the Kindle Fire—one of the highest-profile, most successful Android tablets, even if our reviews of them have always been middling—can't make enough to justify continuing on, that's probably a bad sign for anyone else who's looking to break in. One reason that SoCs have gotten so much faster so quickly is because the competition has been fierce and the potential rewards have been big. For now, this continues to be true—let's hope it stays that way.

Via AnandTech

-----

Today Intel made a sobering, but not entirely unexpected announcement: over the next 3 years Intel will be ramping down its own desktop motherboard business. Intel will continue to supply desktop chipsets for use by 3rd party motherboard manufacturers like ASUS, ASRock and Gigabyte, but after 2013 it will no longer produce and sell its own desktop mITX/mATX/ATX designs in the channel. We will see Haswell motherboards from the group, but that will be the last official hurrah. Intel will stop developing desktop motherboards once the Haswell launch is completed. All Intel boards, including upcoming Haswell motherboards, will carry a full warranty and will be supported by Intel during that period.

This isn't a workforce reduction. Most of the folks who worked in Intel's surprisingly small desktop motherboard division will move on to other groups within Intel that can use their talents. Intel's recently announced NUC will have a roadmap going forward, and some of the desktop board folks will move over there. Intel will continue to produce barebones motherboards for its NUC and future versions of the platform.

Intel will also continue to produce its own form factor reference designs (FFRDs) for Ultrabooks and tablets, which will be where many of these employees will end up as well. As of late Intel has grown quite fond of its FFRD programs, allowing it a small taste of vertical integration (and the benefits that go along with it) without completely alienating its partners. This won't be a transfer of talent to work on smartphone FFRDs at this time however.

The group within Intel responsible for building reference designs that are used internally for testing as well as end up as the base for many 3rd party motherboards will not be impacted by this decision either. The reference board group will continue to operate and supply reference designs to Intel partners. This is good news as it means that you shouldn't see a reduction in quality of what's out there.

It's not too tough to understand why Intel would want to wind down its desktop motherboard business. Intel has two options to keep Wall Street happy: ship tons of product with huge margins and/or generate additional profit (at forgiveably lower margins) that's not directly tied to the PC industry. The overwhelming majority of Intel's business is in the former group. The desktop motherboards division doesn't exactly fit within that category. Motherboards aren't good high margin products, which makes the fact that Intel kept its desktop board business around this long very impressive. Intel doesn't usually keep drains on margins around for too long (look how quickly Intel exited the de-emphasized its consumer SSD business).

The desktop motherboard business lasted so long as a way to ensure that Intel CPUs had a good, stable home (you can't sell CPUs if motherboard quality is questionable). While there was a need for Intel to build motherboards and reference designs 15 years ago, today what comes out of Taiwan is really quite good. Intel's constant integration of components onto the CPU and the resulting consolidation in the motherboard industry has helped ensure that board quality went up.

There's also the obvious motivation: the desktop PC business isn't exactly booming. Late last year word spread of Intel's plans for making Broadwell (14nm Core microprocessor in 2014) BGA-only. While we'll continue to see socketed CPUs beyond that, the cadence will be slower than what we're used to. The focus going forward will be on highly integrated designs, even for the desktop (think all-in-ones, thin mini-ITX, NUC, etc...). Couple that reality with low board margins and exiting the desktop motherboard business  all of the sudden doesn't sound like a bad idea for Intel. 

In the near term, this is probably good for the remaining Taiwanese motherboard manufacturers. They lose a very competent competitor, although not a particularly fierce one. In the long run, it does highlight the importance of having a business not completely tied to desktop PC motherboard sales.

Via ZDNet

-----

Continuing our look in more detail at the 2013 predictions from analyst Mark Anderson, the question of processors and Intel vs ARM is an interesting proxy for the bigger question of "what do we want a computer to be?"

Despite the successes of three generations of Core processors, Intel has plenty of problems to tackle.

Anderson put it bluntly and raised questions about both the business and the technology of CPUs ('CarryAlong' is his term for tablets, phablets, netbooks and other devices that are portable enough to carry and use all day):

"Intel: Long Live the King, the King is Dead. The chip royalty ladder is flipped, as Intel becomes increasingly irrelevant in the world of general computing, while CarryAlong and mobile chipmakers (led by Qualcomm and ARM) are the new William and Kate. For most observers, Intel in 2013 is a component supplier for servers. The best way out of this cul de sac: a new CEO with tech chops."

Quantum barrier

Intel is unbeatable in the laptop space today, but the combination of the popularity of tablets and the laws of physics at 19nm scale and below makes you wonder where they will be in five to 10 years' time. I've wondered if Intel started favouring operations over chip technologists for leadership when they noticed that they were hitting the wall on Moore's Law and had run through Dennard Scaling so much faster than predicted.

(Moore's Law, of course, relates to the number of transistors on a wafer and the cost of the fab to put them there, while Dennard Scaling has to do with keeping the power efficiency of those transistors as good as the previous generation.)

The hi-K metal gates introduced in Sandy Bridge and the tri-gate derivative in Ivy Bridge are making a huge difference in today's Intel chips, but it took 10 years to get that into production. Moreover, a new material is needed to help deal with the fundamental issue of not having enough electrons to get a clean signal to switch the transistor between one and zero without quantum-level effects weirding out the transistor - and if there is such a material, it's still in the labs.

Intel needs to optimise the processes now to make enough time to do the next piece of fundamental science; however, as you hit the quantum level, that fundamental science is a lot harder.

In the meantime I see Intel deliberately continuing to cannibalise itself with Atom because that's preferable to being cannibalised by ARM - and at least uses those fabs that were so expensive when you built them a few years back when they were state of the art. I also see it missing out on LTE, still, which has to hurt its ability to compete in the smartphone market. And if rumours that Haswell chips don't launch until the middle of 2013 are true, Intel would be about six months behind on its usual tick-tock cadence of shrinking a die one year and rolling out a new architecture 12 months later.

Could Intel knock the internal politics on the head while it's at it? I don't understand why CloverTrail isn't yet available in large numbers, but the battle between the Silicon Valley and Israeli chip development teams on direction could be one reason why we don't yet have a Core chip that can support Connected Standby (CS) in Windows 8, even though the power levels should be low enough to make that work.

Connected Standby, Ultrabooks and GPUs

Look at the 17W TDP (thermal design point) of some Ivy Bridge Core models, the promised 15W TDP of Shark Bay Ultrabook processors and then forward to the 8-10W TDP of Haswell, when we might finally get on-package CMOS voltage regulators and support for S0ix active idle - and that's what you need for Connected Standby.

To be fair, it's possible that the reason we don't have Core chips with CS already is that it requires everything to go to a low power fast wake state, not just the CPU - and that's easiest to do when you control all the support components by building an SoC; System on Chip by definition means integration and less variation to deal with. (Qualcomm's huge success has been about not just building a good ARM chip but building it into an integrated platform that can be used to churn out dozens or hundreds of smartphones in a year.)

The Ultrabook programme also gives Intel a way to kick OEMs into line to make them use low-power components (maybe even screens with enough attached RAM to cache what you're looking at on screen when it's not changing fast enough for the GPU to need to be awake), although that's going to face the usual resistance from OEMs who want the choice to carry on with their race to the bottom on $200 tablets and $300 PCs.

Meanwhile, there's Intel's continuing inability to understand that producing graphics drivers that are robust enough to reliably do hardware acceleration is crucial. Notice how few Intel GPUs have been certified for hardware acceleration by Adobe or compare the hardware acceleration you can get in IE 9 and 10 with a 'real' GPU compared to Intel integrated graphics. Offload to GPU is one way to get around Moore's Law (a GPU has far more cores than even the most multi-core of CPUs and they're simplistic and only applicable to easily parallelised problems, but offloading to special purpose hardware is another way ARM systems punch above their weight).

Intel is stuck between the rock of physics, the hard place of mobile computing, the Scylla of expensive fabs and the Charybdis of business models.

Maybe that's why it's trying so hard to become a software company with services for developers. It's been adding NUI concepts like voice recognition and gestures as well as location services under the Computer Continuum umbrella to the language support and development tools it's long offered, and putting its own virtualisation layer under any operating system running on an Intel PC. And, yes, all that does put Intel into competition with an even wider range of industry players...

AMD takes ARMs against the sea of Intel

Inside ARM: The British success story taking the chip world by storm

Via Slash Gear

-----

IBM has developed a light-based data transfer system delivering more than 25Gbps per channel, opening the door to chip-dense slabs of processing power that could speed up server performance, the internet, and more. The company’s research into silicon integrated nanophotonics addresses concerns that interconnects between increasingly powerful computers, such as mainframe servers, are unable to keep up with the speeds of the computers themselves. Instead of copper or even optical cables, IBM envisages on-chip optical routing, where light blasts data between dense, multi-layer computing hubs.

“This future 3D-integated chip consists of several layers connected with each other with very dense and small pitch interlayer vias. The lower layer is a processor itself with many hundreds of individual cores. Memory layer (or layers) are bonded on top to provide fast access to local caches. On top of the stack is the Photonic layer with many thousands of individual optical devices (modulators, detectors, switches) as well as analogue electrical circuits (amplifiers, drivers, latches, etc.). The key role of a photonic layer is not only to provide point-to-point broad bandwidth optical link between different cores and/or the off-chip traffic, but also to route this traffic with an array of nanophotonic switches. Hence it is named Intra-chip optical network (ICON)” IBM

Optical interconnects are increasingly being used to link different server nodes, but by bringing the individual nodes into a single stack the delays involved in communication could be pared back even further. Off-chip optical communications would also be supported, to link the data-rich hubs together.

Although the photonics system would be considerably faster than existing links – it supports multiplexing, joining multiple 25Gbps+ connections into one cable thanks to light wavelength splitting – IBM says it would also be cheaper thanks to straightforward manufacturing integration:

“By adding a few processing modules into a high-performance 90nm CMOS fabrication line, a variety of silicon nanophotonics components such as wavelength division multiplexers (WDM), modulators, and detectors are integrated side-by-side with a CMOS electrical circuitry. As a result, single-chip optical communications transceivers can be manufactured in a conventional semiconductor foundry, providing significant cost reduction over traditional approaches” IBM

Technologies like the co-developed Thunderbolt from Intel and Apple have promised affordable light-based computing connections, but so far rely on more traditional copper-based links with optical versions further down the line. IBM says its system is “primed for commercial development” though warns it may take a few years before products could actually go on sale.

Via The New York Times

-----

SAN FRANCISCO — I.B.M. scientists are reporting progress in a chip-making technology that is likely to ensure that the basic digital switch at the heart of modern microchips will continue to shrink for more than a decade.

The advance, first described in the journal Nature Nanotechnology on Sunday, is based on carbon nanotubes — exotic molecules that have long held out promise as an alternative to silicon from which to create the tiny logic gates now used by the billions to create microprocessors and memory chips.

The I.B.M. scientists at the T.J. Watson Research Center in Yorktown Heights, N.Y., have been able to pattern an array of carbon nanotubes on the surface of a silicon wafer and use them to build hybrid chips with more than 10,000 working transistors.

Against all expectations, silicon-based chips have continued to improve in speed and capacity for the last five decades. In recent years, however, there has been growing uncertainty about whether the technology would continue to improve.

A failure to increase performance would inevitably stall a growing array of industries that have fed off the falling cost of computer chips.

Chip makers have routinely doubled the number of transistors that can be etched on the surface of silicon wafers by shrinking the size of the tiny switches that store and route the ones and zeros that are processed by digital computers.

The switches are rapidly approaching dimensions that can be measured in terms of the widths of just a few atoms.

The process known as Moore’s Law was named after Gordon Moore, a co-founder of Intel, who in 1965 noted that the industry was doubling the number of transistors it could build on a single chip at routine intervals of about two years.

To maintain that rate of progress, semiconductor engineers have had to consistently perfect a range of related manufacturing systems and materials that continue to perform at evermore Lilliputian scale.

The I.B.M. advance is significant, scientists said, because the chip-making industry has not yet found a way forward beyond the next two or three generations of silicon.

“This is terrific. I’m really excited about this,” said Subhasish Mitra, an electrical engineering professor at Stanford who specializes in carbon nanotube materials.

The promise of the new materials is twofold, he said: carbon nanotubes will allow chip makers to build smaller transistors while also probably increasing the speed at which they can be turned on and off.

In recent years, while chip makers have continued to double the number of transistors on chips, their performance, measured as “clock speed,” has largely stalled.

This has required the computer industry to change its designs and begin building more so-called parallel computers. Today, even smartphone microprocessors come with as many as four processors, or “cores,” which are used to break up tasks so they can be processed simultaneously.

I.B.M. scientists say they believe that once they have perfected the use of carbon nanotubes — sometime after the end of this decade — it will be possible to sharply increase the speed of chips while continuing to sharply increase the number of transistors.

This year, I.B.M. researchers published a separate paper describing the speedup made possible by carbon nanotubes.

“These devices outperformed any other switches made from any other material,” said Supratik Guha, director of physical sciences at I.B.M.’s Yorktown Heights research center. “We had suspected this all along, and our device physicists had simulated this, and they showed that we would see a factor of five or more performance improvement over conventional silicon devices.”

Carbon nanotubes are one of three promising technologies engineers hope will be perfected in time to keep the industry on its Moore’s Law pace.
Graphene is another promising material that is being explored, as well as a variant of the standard silicon transistor known as a tunneling field-effect transistor.

Dr. Guha, however, said carbon nanotube materials had more promising performance characteristics and that I.B.M. physicists and chemists had perfected a range of “tricks” to ease the manufacturing process.

Carbon nanotubes are essentially single sheets of carbon rolled into tubes. In the Nature Nanotechnology paper, the I.B.M. researchers described how they were able to place ultrasmall rectangles of the material in regular arrays by placing them in a soapy mixture to make them soluble in water. They used a process they described as “chemical self-assembly” to create patterned arrays in which nanotubes stick in some areas of the surface while leaving other areas untouched.

Perfecting the process will require a more highly purified form of the carbon nanotube material, Dr. Guha said, explaining that less pure forms are metallic and are not good semiconductors.

Dr. Guha said that in the 1940s scientists at Bell Labs had discovered ways to purify germanium, a metal in the carbon group that is chemically similar to silicon, to make the first transistors. He said he was confident that I.B.M. scientists would be able to make 99.99 percent pure carbon nanotubes in the future.

Via Venture Beat

-----

IBM has been shipping computers for more than 65 years, and it is finally on the verge of creating a true electronic brain.

Big Blue is announcing today that it, along with four universities and the Defense Advanced Research Projects Agency (DARPA), have created the basic design of an experimental computer chip that emulates the way the brain processes information.

IBM’s so-called cognitive computing chips could one day simulate and emulate the brain’s ability to sense, perceive, interact and recognize — all tasks that humans can currently do much better than computers can.

Dharmendra Modha (pictured below right) is the principal investigator of the DARPA project, called Synapse (Systems of Neuromorphic Adaptive Plastic Scalable Electronics, or SyNAPSE). He is also a researcher at the IBM Almaden Research Center in San Jose, Calif.

“This is the seed for a new generation of computers, using a combination of supercomputing, neuroscience, and nanotechnology,” Modha said in an interview with VentureBeat. ”The computers we have today are more like calculators. We want to make something like the brain. It is a sharp departure from the past.”

If it eventually leads to commercial brain-like chips, the project could turn computing on its head, overturning the conventional style of computing that has ruled since the dawn of the information age and replacing it with something that is much more like a thinking artificial brain. The eventual applications could have a huge impact on business, science and government. The idea is to create computers that are better at handling real-world sensory problems than today’s computers can. IBM could also build a better Watson, the computer that became the world champion at the game show Jeopardy earlier this year.

We wrote about the project when IBM announced the project in November, 2008 and again when it hit its first milestone in November, 2009. Now the researchers have completed phase one of the project, which was to design a fundamental computing unit that could be replicated over and over to form the building blocks of an actual brain-like computer.

Richard Doherty, an analyst at the Envisioneering Group, has been briefed on the project and he said there is “nothing even close” to the level of sophistication in cognitive computing as this project.

This new computing unit, or core, is analogous to the brain. It has “neurons,” or digital processors that compute information. It has “synapses” which are the foundation of learning and memory. And it has “axons,” or data pathways that connect the tissue of the computer.

While it sounds simple enough, the computing unit is radically different from the way most computers operate today. Modern computers are based on the von Neumann architecture, named after computing pioneer John von Neumann and his work from the 1940s.

In von Neumann machines, memory and processor are separated and linked via a data pathway known as a bus. Over the past 65 years, von Neumann machines have gotten faster by sending more and more data at higher speeds across the bus, as processor and memory interact. But the speed of a computer is often limited by the capacity of that bus, leading some computer scientists to call it the “von Neumann bottleneck.”

With the human brain, the memory is located with the processor (at least, that’s how it appears, based on our current understanding of what is admittedly a still-mysterious three pounds of meat in our heads).

The brain-like processors with integrated memory don’t operate fast at all, sending data at a mere 10 hertz, or far slower than the 5 gigahertz computer processors of today. But the human brain does an awful lot of work in parallel, sending signals out in all directions and getting the brain’s neurons to work simultaneously. Because the brain has more than 10 billion neuron and 10 trillion connections (synapses) between those neurons, that amounts to an enormous amount of computing power.

IBM wants to emulate that architecture with its new chips.

“We are now doing a new architecture,” Modha said. “It departs from von Neumann in variety of ways.”

The research team has built its first brain-like computing units, with 256 neurons, an array of 256 by 256 (or a total of 65,536) synapses, and 256 axons. (A second chip had 262,144 synapses) In other words, it has the basic building block of processor, memory, and communications. This unit, or core, can be built with just a few million transistors (some of today’s fastest microchips can be built with billions of transistors).

Modha said that this new kind of computing will likely complement, rather than replace, von Neumann machines, which have become good at solving problems involving math, serial processing, and business computations. The disadvantage is that those machines aren’t scaling up to handle big problems well any more. They are using too much power and are harder to program.

The more powerful a computer gets, the more power it consumes, and manufacturing requires extremely precise and expensive technologies. And the more components are crammed together onto a single chip, the more they “leak” power, even in stand-by mode. So they are not so easily turned off to save power.

The advantage of the human brain is that it operates on very low power and it can essentially turn off parts of the brain when they aren’t in use.

These new chips won’t be programmed in the traditional way. Cognitive computers are expected to learn through experiences, find correlations, create hypotheses, remember, and learn from the outcomes. They mimic the brain’s “structural and synaptic plasticity.” The processing is distributed and parallel, not centralized and serial.

With no set programming, the computing cores that the researchers have built can mimic the event-driven brain, which wakes up to perform a task.

Modha said the cognitive chips could get by with far less power consumption than conventional chips.

The so-called “neurosynaptic computing chips” recreate a phenomenon known in the brain as a “spiking” between neurons and synapses. The system can handle complex tasks such as playing a game of Pong, the original computer game from Atari, Modha said.

Two prototype chips have already been fabricated and are being tested. Now the researchers are about to embark on phase two, where they will build a computer. The goal is to create a computer that not only analyzes complex information from multiple senses at once, but also dynamically rewires itself as it interacts with the environment, learning from what happens around it.

The chips themselves have no actual biological pieces. They are fabricated from digital silicon circuits that are inspired by neurobiology. The technology uses 45-nanometer silicon-on-insulator complementary metal oxide semiconductors. In other words, it uses a very conventional chip manufacturing process. One of the cores contains 262,144 programmable synapses, while the other contains 65,536 learning synapses.

Besides playing Pong, the IBM team has tested the chip on solving problems related to navigation, machine vision, pattern recognition, associative memory (where you remember one thing that goes with another thing) and classification.

Eventually, IBM will combine the cores into a full integrated system of hardware and software. IBM wants to build a computer with 10 billion neurons and 100 trillion synapses, Modha said. That’s as powerful than the human brain. The complete system will consume one kilowatt of power and will occupy less than two liters of volume (the size of our brains), Modha predicts. By comparison, today’s fastest IBM supercomputer, Blue Gene, has 147,456 processors, more than 144 terabytes of memory, occupies a huge, air-conditioned cabinet, and consumes more than 2 megawatts of power.

As a hypothetical application, IBM said that a cognitive computer could monitor the world’s water supply via a network of sensors and tiny motors that constantly record and report data such as temperature, pressure, wave height, acoustics, and ocean tide. It could then issue tsunami warnings in case of an earthquake. Or, a grocer stocking shelves could use an instrumented glove that monitors sights, smells, texture and temperature to flag contaminated produce. Or a computer could absorb data and flag unsafe intersections that are prone to traffic accidents. Those tasks are too hard for traditional computers.

Synapse is funded with a $21 million grant from DARPA, and it involve six IBM labs, four universities (Cornell, the University of Wisconsin, University of California at Merced, and Columbia) and a number of government researchers.

For phase 2, IBM is working with a team of researchers that includes Columbia University; Cornell University; University of
California, Merced; and University of Wisconsin, Madison. While this project is new, IBM has been studying brain-like computing as far back as 1956, when it created the world’s first (512 neuron) brain simulation.

“If this works, this is not just a 5 percent leap,” Modha said. “This is a leap of orders of magnitude forward. We have already overcome huge conceptual roadblocks.”

[Photo credits: Dean Takahashi, IBM]

Via Tech Hive

-----

If you follow the world of Android tablets and phones, you may have heard a lot about Tegra 3 over the last year. Nvidia's chip currently powers many of the top Android tablets, and should be found in a few Android smartphones by the end of the year. It may even form the foundation of several upcoming Windows 8 tablets and possibly future phones running Windows Phone 8. So what is the Tegra 3 chip, and why should you care whether or not your phone or tablet is powered by one?

Nvidia's system-on-chip

Tegra is the brand for Nvidia's line of system-on-chip (SoC) products for phones, tablets, media players, automobiles, and so on. What's a system-on-chip? Essentially, it's a single chip that combines all the major functions needed for a complete computing system: CPU cores, graphics, media encoding and decoding, input-output, and even cellular or Wi-Fi communcations and radios. The Tegra series competes with chips like Qualcomm's Snapdragon, Texas Instruments' OMAP, and Samsung's Exynos.

The first Tegra chip was a flop. It was used in very few products, notably the ill-fated Zune HD and Kin smartphones from Microsoft. Tegra 2, an improved dual-core processor, was far more successful but still never featured in enough devices to become a runaway hit.

Tegra 3 has been quite the success so far. It is found in a number of popular Android tablets like the Eee Pad Transformer Prime, and is starting to find its way into high-end phones like the global version of the HTC One X (the North American version uses a dual-core Snapdragon S4 instead, as Tegra 3 had not been qualified to work with LTE modems yet). Expect to see it in more Android phones and tablets internationally this fall.

4 + 1 cores

Tegra 3 is based on the ARM processor design and architecture, as are most phone and tablet chips today. There are many competing ARM-based SoCs, but Tegra 3 was one of the first to include four processor cores. There are now other quad-core SoCs from Texas Instruments and Samsung, but Nvidia's has a unique defining feature: a fifth low-power core.

All five of the processor cores are based on the ARM Cortex-A9 design, but the fifth core is made using a special low-power process that sips battery at low speeds, but doesn't scale up to high speeds very well. It is limited to only 500MHz, while the other cores run up to 1.4GHz (or 1.5GHz in single-core mode).

When your phone or tablet is in sleep mode, or you're just performing very simple operations or using very basic apps, like the music player, Tegra 3 shuts down its four high-power cores and uses only the low-power core. It's hard to say if this makes it far more efficient than other ARM SoCs, but battery life on some Tegra 3 tablets has been quite good.

Good, not great, graphics

Nvidia's heritage is in graphics processors. The company's claim to fame has been its GPUs for traditional laptops, desktops, and servers. You might expect Tegra 3 to have the best graphics processing power of any tablet or phone chip, but that doesn't appear to be the case. Direct graphics comparisons can be difficult, but there's a good case to be made that the A5X processor in the new iPad has a far more powerful graphics processor. Still, Tegra 3 has plenty of graphics power, and Nvidia works closely with game developers to help them optimize their software for the platform. Tegra 3 supports high-res display output (up to 2560 x 1600) and improved video decoding capabilities compared to earlier Tegra chips.

Do you need one?

The million-dollar question is: Does the Tegra 3 chip provide a truly better experience than other SoCs? Do you need four cores, or even "4 + 1"? The answer is no. Most smartphone and tablet apps don't make great use of multiple CPU cores, and making each core faster can often do more for the user experience than adding more cores. That said, you shouldn't avoid a product because it has a Tegra 3 chip, either. Its performance and battery life appear to be quite competitive in today's tablet and phone market. Increasingly, the overall quality of a product is determined by its design, size, weight, display quality, camera quality, and other features more than mere processor performance. Consider PCWorld's review of the North American HTC One X; with the dual-core Snapdragon S4 instead of Tegra 3, performance was still very impressive.