Your browser is not supported. Please update it.

11 October 2012

ARM and Intel move to squeeze PowerPC in networking processor market

Significance: ARM has been creeping up the processor food chain from devices into Intel’s server heartland, and into the network itself. It has been an important element of Texas Instruments’ infrastructure portfolio since that firm broadened from DSPs into full platforms, but only recently has it won a place with PowerPC devotees Freescale and LSI. They have given ARM a role mainly because of its cost and power credentials, but are far from osting Power entirely. Meanwhile, Intel has been building on strengths at the network edge to announce products for high end communications and even for base stations. The first victims of the ARM/Intel pincer movement are likely to be custom designs and ASICs, which become less attractive in the low cost economics of modern mobile and cloud data networks. After that, the big question is whether PowerPC and, in other sectors, MIPS, will be squeezed out, or saved by the rising interest in multi-processor frameworks like Freescale’s Layerscape.

The processor architectures of ARM and Intel are battling on three fronts ‘ mobile devices, servers and embedded chips ‘ and are increasingly moving on the mobile network, where PowerPC and custom chips have ruled the roost. ARM unveiled its CoreLink interconnect technology this week, along with a memory controller, both targeted at networking and storage equipment.

Texas Instruments, LSI and even – to a more limited extent – PowerPC’s greatest backer Freescale, have included ARM cores in their network infrastructure platforms. The UK-based firm sees expanding opportunities as the makers of base station and communication processors rethink their designs in order to cope with the rising demands of mobile data traffic.

At the Linley Tech Processor Conference in San Jose, California ARM said its latest offerings should ship in commercial products in 2014. Its latest intellectual property blocks are the CoreLink CCN-504 interconnect (in effect an on-die network) and the CoreLink DMC-520 dynamic memory controller.

The two designs work together, with the former providing the ‘glue’ to connect cores and cache components, within a chip and also to other system elements. It promises real world throughput of 1Tbps speeds (peaks of 1.5Tbps) with 128-bit internal bus, 16Mbytes of shared L3 cache and dual memory channels to link to the DMC-520 memory controllers. There is support for up to ARM cores, whether the current high end 32-bit Cortex A15, or the upcoming 64-bit ARM v8. 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. The 504 can be linked to several flavors of high speed networking technology, such as 40Gig Ethernet, and is suited to wired and wireless infrastructure.

ARM focuses its differentiation in the usual area, its low power consumption, which is becoming a critical issue for servers and networks, not just devices, as operators require ever more horsepower to manage their data. Its low energy credentials underpin the whole strategy, of which CoreLink is a part, to push its architecture into as many markets as possible, squeezing x86, PowerPC and Mips.

In this case, the target applications are communications and storage at the edge of the network, an area which Intel is also chasing. For both, according to Jim McGregor, principal analyst at Tirias Research, the first wins will come by displacing custom ASICs or specialized RISC architectures, before the two giants really go head-to-head.

As in mobile devices, ARM will stand or fall on its ecosystem and the breadth of support for its platforms, which reduce the risk of relying on a relatively small player, accelerate R&D efforts, and ensure control does not rest with one major firm. In the networking market, SoC supporters include Broadcom, Cavium, Freescale, Huawei, LSI, Mindspeed and TI. Freescale still uses its trademark PowerPC for much of its networking and base station range, but recently adopted ARM for small cells and there is considerable speculation that it will make a fuller conversion over time as power reduction pressures mount.

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, a high profile start-up leading the push to bring low power SoCs to cloud servers. LSI was an important convert to ARM’s network processor efforts at this year’s Mobile World Congress, when the former ‘ like Freescale a PowerPC user ‘ announced it would also license the ARM A15 core for use in its next generation offerings and in ASICs. It expects to use PowerPC and A15 SoCs in parallel, although product line director Tareq Bustami conceded the A15 was ‘more efficient’.

The push to combine the best qualities of different architectures is gaining force, because of the diverse demands of the infrastructure segments ‘ many believe that, in contrast to devices, one processor platform cannot satisfy all needs. ARM itself points out that the networking and server segments are broad, with a far wider range of applications than in devices, hence why many processor architectures and custom designs have survived.

Hence it stresses flexibility, saying its approach allows chip suppliers to create a broad variety of SoCs with different target functions, while maintaining a common base platform. And the new interconnect can link many different elements which could be included in a SoC, such as digital signal processors and vendor-specific functions. However, Intel responds that ARM will never achieve the software compatibility of x86. The company recently unveiled its own latest attack on the carrier network, Crystal Forest.

The firm aims to hit Freescale in two of its strongholds – the data plane, and next year, the base station. It will attempt the first feat with Crystal Forest, which will be in production late this year. It will target existing Intel markets like edge routers and firewalls, but most importantly, it will take the Xeon x86 processor up a level to go head-to-head with Freescale, Broadcom/NetLogic and Cavium at the high end. Crystal Forest pairs a new generation Xeon processor (Sandy Bridge) with a companion chip called Cave Creek, which is now sampling. Cave Creek replaces the usual Xeon server companion with silicon suited to midrange or high end routers ‘ most importantly, hardware accelerators for cryptography, compression and pattern matching. These accelerators are new versions of cores used in previous Intel chips such as Tolapai and IXP and form part of Intel’s QuickAssist platform for efficient packet processing. CrystalForest promises to deliver up to 160m packets per second for Layer 3 packet forwarding, a level which, according to the chip giant, was previously possible only with specialized processors. Intel will also release a set of software libraries and algorithms called the Data Plane Development Kit to help accelerate use of the x86 in high end packet processing, promising a fivefold performance boost in that area.

It is important that the architectures for handling massive amounts of multimedia data are simplified, argues Intel, which says that manufacturers currently need to combine many highly specialized coprocessors with different programming models to meet demands of scalability. This increases costs and complexity, but Crystal Forest promises to consolidate three key communications workloads ‘ application, control and packet processing ‘ on a standard architecture.

The Cave Creek coprocessor could challenge rivals like the Cavium Nitrox, NetLogic XLP or Freescale QorIQ accelerators, especially where OEMs are using x86 already in the control plane, says Linley Group. It will be applied to other segments too, and can be paired with less hefty CPUs than Xeon. For instance, it can be coupled with notebook chips such as the Core i3, i5 and i7 to provide a low power communications platform suited to an edge router or appliance.

And next year is likely to see a Xeon-based platform with a companion chip targeted at digital signal processing in base stations, again hitting at a Freescale heartland, as well as at Texas Instruments. This will come with a set of signal processing libraries, currently under development, for handling Layer 1 baseband processing (though base stations will still need some external parts such as Viterbi decoders).

Intel may be majoring on software harmonization and simplifying the architecture, and ARM on power and flexibility, but PowerPC has its own advantages, notably its incumbent position. While several of its devotees are converting to ARM, they generally want the best of both worlds, so the architecture remains central to the roadmaps of Freescale and LSI, at least for now. And the next step is to support processor-neutral architectures ‘ seen in AMD’s Heterogeneous System Architecture and Freescale’s Layerscape. This should allow firms to move on from supporting diverse customer needs with entirely separate product ranges, based on different processors, and evolve a truly multi-platform strategy.

Layerscape will provide a modular architecture able to support different cores (ARM, PowerPC and accelerators, at least at first). And all of these will be programmable in standard, high level languages such as C, easing the task of developers even on multicore designs. Software for Freescale’s QorIQ chips is currently programmed in assembly languages, which is labor intensive and often customized. Ease of programmability is a significant factor when companies are choosing a processor architecture, and has often stood Intel in good stead when driving x86 into communications applications.

The first two families of Layerscape chips will sample in mid-2013 when the new software programming environment should be complete, Freescale told EETimes this summer. Tom Dietrich, general manager of Freescale’s Networking & Multimedia Solutions group, said when unveiling the new platform: “Layerscape really will set the benchmark for the next 10 years of network processing applications …. with a software-aware framework”. This has three independent, scalable layers ‘ one for general purpose processing; an accelerated packet processing layer; and an express packet I/O layer. This modular approach will enable Freescale to design QorIQ devices with expanded, reduced or removed layers as needed, depending on the application.

As in most elements of the wireless ecosystem, suppliers are keen to keep their options open and avoid handing too much power to a single approach. Genuine multi-platform architectures are in their infancy but suggest that ARM and Intel will have to resign themselves to coexisting in many customers’ roadmaps in the server and networking spaces. This will not be a head-to-head, with outright victory as the stakes, unlike the device market. The major question marks hang over PowerPC, and how strong its role will be in future. The early flexible frameworks like Layerscape appear to throw it a lifeline by avoiding the need for chip suppliers to make an either/or choice. Given the new economics of the mobile data network and the cloud, ARM and Intel can certainly look to benefit from a shift away from custom designs and towards standardized platforms and programming tools. However, they may also need to accept that their architectures will not be able, alone, to support all the daunting requirements of the modern data network.