Your browser is not supported. Please update it.

1 November 2012

ARM shows the future face of smartphones and servers

Significance: ARM has been gradually putting together the key pieces of a strategy that will expand its famous mobile processor architecture across every link in the cloud-connected chain, from servers and base stations to the internet of things. In unveiling its processor collection for 2014, it slotted in the final piece of the jigsaw ‘ two Cortex-A processor cores, designed to take advantage of all its recent strategic moves, including 64-bit support, the big.LITTLE flexible multicore idea, and the CoreLink fabric to enable massively parallel designs. Both ARM and Intel are seeing their platforms for mobile products, PCs and servers gaining more commonality. Smartphones are already multicore and mobile gadgets could require massively parallel capabilities in 5-10 years, especially if the burden of web service processing shifts from the server to the client, as Intel believes. And the servers can be built, now that ARM has a 64-bit architecture, using huge numbers of cellphone chips. This overreaching ARM architecture is now largely in place, and set for commercial products in 2014. Now it passes the baton to its licensees, including new recruit AMD, to create the processors that will turn the vision to reality.

ARM showed off what it hopes will be the shape of 2014 in electronic devices from smartphones to servers to base stations. Perhaps betraying its understated British roots, the Cambridge-based firm was down to earth about its latest generation of processor cores. While its key smartphone customers may present a minor performance tweak as a ‘revolution’, you had to read ARM’s statements carefully to realize that what they are attempting actually does class as revolutionary. The Cortex-A50 brings together several radical changes which the company has been introducing over the past year or more, as it aims to be the dominant architecture not just in its mobile stronghold but in servers, network infrastructure and the Internet of Things. Converging in these products are the move to 64-bit platforms with ARMv8; support for new manufacturing processes, initially 20nm; the CoreLink fabric for interconnecting mixed cores in parallel arrays; and the important big.LITTLE architecture, which aims to achieve a perfect ratio between performance and power efficiency.

The first Cortex-A50 designs, the A57 and A53, are targeted at traditional ARM targets ‘ superphones and tablets for the former, entry level smartphones for the latter. But they both come in 64-bit and low power 32-bit variants, and can be harnessed together in the big.LITTLE arrangement to enable multicore devices which balance tasks between processors with different strengths (horsepower for compute intensive activities, low energy for communications, for instance).

Big.LITTLE was introduced last year for the A15 and A7 cores. The low power, low cost A7 can be used alongside the high performance A15 as part of a heterogeneous multicore strategy, in which cores can be switched on and off to conserve power. The smaller chip supports the same virtualization and extended addressing of the A15, so that individual cores can be selected to run different applications based on power efficiency needs. Basic and always-on tasks run on one or more A7 cores, while processor intensive tasks move to the A15 cores. This goes further than current designs which allocated activities to different cores, as these have been confined to offloading specific work such as graphics to specialist cores. In ARM’s system, the dynamic selection of generic cores is invisible to the applications and middleware, and is managed by ARM system IP. The transfer of tasks between cores is triggered by the same mechanism which handles dynamic voltage and frequency scaling, now common in high end chips.

ARM expects A15/A7 combinations to appear in commercial products next year, mainly smartphones and embedded systems (Freescale has already demonstrated big.LITTLE in the latter category). Then the A57/A53 combination will take over in 2014, and extend the concept into servers and other platforms. The A57 claims to deliver three times the performance of current superphone processors at the same power consumption, while the A53 claims four times the energy efficiency and greater horsepower than current mainstream smartphone apps chips. The A53 is also the world’s smallest 64-bit processor, says ARM, while the A57 ‘provides computer performance comparable to a legacy PC, while operating in a mobile power budget’.

Further family members will support massively parallel arrays, for use in low energy cloud servers, as well as configurations optimized for communications equipment, consumer electronics gadgets, home kit such as set-top boxes, and other form factors. The UK firm expects to see a wide range of products coming to market in 2014 based on a broadened A50 platform, including tablets, power efficient cloud servers, network infrastructure such as base station chips and communications processors, and a new generation of consumer electronics devices.

By 2014, of course, a whole new generation of devices will have come and gone, and there will be new form factors and user experiences to support. ARM hopes the flexibility of its evolving architecture, and the ability to scale from a single low power processor to hundreds with mixed abilities, will adapt to whatever the gadget designers throw at the industry amid the rapid change of the post-PC world. And of course, the move to 64-bit is seen as essential to keep up with performance demands across most device types, and to make a serious bid for the enterprise and carrier systems.

In all these areas, ARM will primarily take on Intel. MIPS remains a strong architecture in set-top boxes and home gateways, while PowerPC is on the defensive, but still entrenched, in network equipment. In the latter’s case, Freescale’s recent decision to introduce ARM into its PowerPC-based base station line-up, at least at the low end, was taken as a sign that ARM had already won the battle for the energy efficient side of this sector, and would rise on the tide of small cell adoption in future. However, some disagree. There is even a new start-up, Synergy, targeting the low power cloud server space with PowerPC, refusing to take the common line that the architecture will not penetrate any new sectors but will need to maximize its presence in existing strongholds. Texas-based Synergy will be the only firm apart from PowerPC co-creator IBM to use the architecture in servers. It is led by Bill Mapp, a former principal in IBM’s services group,

Flexible architectures which can support a mixture of processors, as seen at AMD and Freescale, offer a way to prolong the relevance of these platforms, but they are unlikely to expand significantly into new sectors. If we assume, like many analysts, that the manufacturers will increasingly demand a common architecture for every kind of web-enabled product, this will certainly be a two-headed battle between ARM and Intel. That, in itself, is a big assumption ‘ while it is almost inconceivable that Intel will embrace ARM, as it once did, many of its rivacl chipmakers, including Broadcom and Freescale, differentiate themselves by relying on multiple platforms (for example, Broadcom’s MIPS-based home gear and Freescale’s Power PC communications products).

But ARM and Intel will be the only majors realistically targeting every segment. To achieve the revenues and growth it needs to stay on top of the market, it is essential that ‘Intel Inside’ becomes ‘Intel Everywhere’, and these days, that challenge relies entirely on the x86 architecture. Like ARM, Intel is expanding from its heartlands ‘ in servers and PCs ‘ into more hostile territory (mobile devices, embedded products and communications gear). As ARM has spent many years mounting a credible challenge to x86 in performance terms, so Intel has been through the same painful process to establish its low energy credentials. The perfect balance of the two ‘ which will vary in different products ‘ is the goal of both.

Though ARM tends to attract media favour for being the plucky challenger, it probably has an even higher mountain to climb in servers than Intel does in mobile devices, where the world’s largest chipmaker has finally seen a small but significant breakthrough in smartphones ‘ and has the home advantage when it comes to the most direct head-to-head with ARM so far, in Windows 8. Here, the choice between energy efficient, fully mobile systems and high performance products will be most stark and the consumers will, for the first time, be able to make that choice within a single category of gadgets (tablets, ultrabooks and hybrids).

The outcome is likely to show that the market still needs both x86 and ARM. Both have made progress in addressing their weaknesses but are still perceived by OEMs to have contrasting capabilities. The next step will be to beat one another at their own game. To be a real challenger in servers – rather than an interesting alternative focused only on products where energy efficient is paramount ‘ ARM must outdo Intel in performance and software base too, just as Intel must do the same in power consumption.

Its success has been driven by the changes in computing devices ‘ compute power coming into previously dumb gadgets like phones, set-top boxes and music players; always-on web connectivity being added to standalone products like fridges and PCs; massive cloud servers looking for a more cost effective and energy efficient solution to supporting all those connected devices. But while it is essential to play the low power card in the server rack market, ARM risks doing this at the expense of performance. At the launch of big.LITTLE last year, Tom Cronk, deputy chief of the processor division, said: “We took the A15 to market last year because we needed to push the performance envelope. But power efficiency is the most important thing for ARM. The A7 is the most efficient core yet.’

That needs to change. The A57 and its high performance successors aim to do that, though we need to see even more powerful variants, optimized for servers, to be fully convinced. But the 64-bit Cortex designs do finally achieve what the hype has promised for so long ‘ a cellphone chip in a high end server. Combined with the big.LITTLE approach and the CoreLink fabric, a robust story for the high end space is emerging.

Both the new cores are supported by the ARM CoreLink 400 and new CoreLink 500 series system IP fabric solutions, and partners can scale up to massively parallel enterprise platforms like those touted by Calxeda and others. The CoreLink interconnect technology is important because it allows many combinations of cores and other system elements to be integrated in a scalable way, up to massively parallel arrays.

For instance, ARM last month showed off CoreLink with a memory controller targeted at networking and storage equipment. It unveiled two intellectual property blocks, the CoreLink CCN-504 interconnect (in effect an on-die network) and the CoreLink DMC-520 dynamic memory controller. They work together, with the former providing the ‘glue’ to connect cores and cache components, within a chip and also to other system elements. The CCN-504 can scale up from single-core CPUs to four clusters of four cores each, meaning the common architecture can power infrastructure from small cells to macrocells to communications.

In the case of the interconnect, LSI has taken the role of lead partner, with early access to the technology. It will start shipping sample SOCs next year. The other first stage supporter is Calxeda, the high profile start-up leading the push to bring low power SoCs to cloud servers.

For mobile devices, integration with different elements is required, notably graphics or video processors which can offload those intensive tasks. The new A50 series chips offer interoperability with ARM’s Mali graphics processor family and CoreLink supports clusters of mixed Cortex-A and Mali cores.

The clearest sign that it is convincing the server market is the belated conversion of AMD, the other main x86 supporter. The initial public licensees for the A50 Series include AMD and the pioneer of the low power server category, Calxeda. Others include Broadcom, HiSilicon, Samsung and STMicro.

Announcing the cooperative project between ARM and AMD, the latter’s SVP for global business units, Lisa Su, said: ‘ARM’s Cortex-A50 series processors, combined with AMD’s unique low power dense compute fabric, will bring new levels of choice and energy efficiency to the data center compute landscape and ignite a new 64-bit server ecosystem.’

AMD will start offering chips based on ARMv8 in 2014, said CEO Rory Read, and will also combine these processors with designs from start-up SeaMicro, which the x86 firm acquired last year. AMD has belatedly woken up to the need to change its lifelong strategy of providing a direct alternative to Intel in classic x86 categories. It has not followed the leader into mobile chips, and its market share has fallen to under 5% in server processors. It now needs to differentiate itself from Intel, and that means adopting ARM, with its huge ecosystem, in order to target emerging new server sectors where it has a chance of outwitting its main competitor. ARM has often said it would be a thankless task to try to oust Intel from the conventional enterprise server ‘ instead it wants to push its architecture at the very different needs of the next generation server, notably the requirement to reduce power consumption when cloud-based businesses are running on thousands of servers at once.

Read said AMD will continue to create custom x86 processor cores for a long time and many expect it to take an architectural licence from ARM, which would enable it to create custom cores in that market too. However, the firm has many challenges ahead ‘ being the only company to support both x86 and ARM involves hefty resources and difficult balancing acts, as Intel itself found out when it acquired the StrongARM business. AMD has a good vision of multi-architecture fabrics and a growth in custom revenues built around that, but in the short term it has no mobile offering below slim notebooks, and it will have to wait almost two years to have a platform for the low energy server space ‘ a period in which others may leap ahead, particularly if a major rival snaps up the innovative Calxeda.

Currently, at least five companies are designing ARM-based server SoCs, including Calxeda; Applied Micro, another defector from PowerPC; Marvell with Armada XP; Cavium; and reportedly Samsung.

Calxeda itself set out its 64-bit strategy in the wake of ARM’s announcements. It will release a 32-bit server platform, Midway, in 2013 and move to 64-bit in 2014. The former will use an updated version of the firm’s on-chip Fabric Switch to support dynamic power and routing optimization for clusters of servers. Midway will have a “comparable performance per watt” to Calxeda’s current generation of ARM-based 5W EnergyCard servers. The 64-bit design, Lago, will come with a further upgrade to Fabric Switch, meaning the server will be able to connect “hundreds of thousands of nodes” together, the company said.

Also in the high end space, Nvidia will help take ARM into new markets. The graphics processor specialist has, over the past couple of years, made an impact on apps processors too, via its ARM-based Tegra platform for smartphones and tablets. Now it wants to spread its wings beyond the GPU in servers and supercomputers too, and plans to extend its Tesla high end GPU range with a combined CPU/GPU offering. Although timescales are vague, this could be a powerful challenger to AMD’s similarly integrated architecture, while Nvidia also aims to introduce powerful GPUs into the mobile Tegra offering.

“Tegra is going to become GPU computing capable in the not-so-distant future. Sometime this decade we are also going to start bringing integrated CPUs and GPUs together in the Tesla line,” said CTO Steve Scott in an interview reported by Techworld. Nvidia is already working on chips using ARMv8. “We are working on Project Denver, which aims to make a very high performance ARM core based on ARM64 that can be used in the range of Nvidia products,” the firm said. It has already showcased the potential of ARM in massively parallel computing, demonstrating a prototype supercomputer last year based on 1,000 quad-core Tegra 3 chips.

So what about Intel’s response? The firm is working on its own massively multicore architecture, MIC (Many Integrated Cores) and on future Xeon x86 processors with 48 cores. It also believes this architecture will be sufficiently compact and low power to appear in mobile devices ‘ in a 5-10 year time span. A late convert to multicore ‘ Intel’s traditional strength was in cranking up a single processor ‘ it now believes the computing demands of small devices will one day justify massively parallel mobile gadgets, and wants to be ready when that trend materializes.

Researchers at Intel Labs in Barcelona have created a prototype of a 48-core chip for smartphones and tablets, which can run different apps or tasks on different cores at the same time.

CTO Justin Rattner envisages a reversal of current thinking that the cloud servers will do all the legwork while the actual devices become smaller and less powerful. In fact, after a phase of thin clients, he believes it will become ‘just not practical to just take sound and pictures and send them up to the cloud and expect that some server is going to perform those tasks. So a lot of that will be pushed out to the client devices.’ While current applications look to reduce strain on small gadgets, there is a performance trade-off, as the rising complaints about slow response from Apple’s Siri or iCloud services indicate. They may be viable for relatively low performance handsets, but they will stretch even an LTE network with their constant interactions. In an interview with IDG, Rattner identified several applications which he believes will be mainly suited to mobile usage, but will require hugely parallel devices – encoding massive videos, always-on augmented reality, and large-scale multitasking were among them.

Of course, it would be in Intel’s interests to see a world where its high performance talents were more vital to the mobile world, while ARM would prefer a server-oriented view which would stimulate demand for its low energy designs in an untapped sectors. But despite significant progress for the ARM platform at the high end, the volume in that market will stay with mobile and embedded devices for years to come, and the UK firm itself is acutely aware of the need to play on all fronts ‘ something AMD has not achieved, even as other players from the PC space, such as Nvidia, have stretched their horizons considerably on the back of an ARM licence. AMD, then, may turn out to be a symbolic win for ARM rather than a huge contributor to its growth. The key to ARM’s expansion is not servers or even devices, but the cloud. As the comments from Intel’s Rattner highlight, nobody really knows how the processing power will be shared in future between client and server. The important thing for both Intel and ARM is to be well positioned for either outcome. With the new processor family and the addition of more and more surrounding technologies like CoreLink and Mali, ARM has certainly grown up and has a platform ready to compete across the gamut of cloud-connected systems. Whether all this really will lead to a redefinition of the smart device and the server in 2014 now depends on the innovations of its partners, and how effectively Intel responds.