Significance: Intel and Texas Instruments are now rivals for the smartphone processor, but they may face an even more intense battle on the infrastructure side. Once, Intel’s server markets and TI’s base stations would hardly have intersected, but with the convergence of computing and communications, these products ‘ and the network processors in between ‘ are getting closer together. That is seeing TI pushing its Keystone architecture into other network elements and even into the cloud server, while Intel is adapting its Xeon platform for Cloud-RANs, small cells and even conventional base stations.
Base stations and servers have always inhabited entirely different landscapes, but as the lines between communications and computing blur, they are moving closer together. Both have to handle rapidly increasing loads of multimedia data; both have to operate at lower power levels; and through trends like Cloud RAN, the server is finding its way into the heart of the telco network.
No wonder, then, that chip providers from these very different markets should be looking over the fence into each other’s gardens. Intel is a great supporter of Cloud RAN, which virtualizes the activities of many base stations on a centralized server, and has also infiltrated Atom and Xeon processors into ‘companions’ for small cells. Next year it promises to push Xeon into base stations, moving into the heartland of TI and Freescale. Conversely, those two firms are extending their architectures into every element of the carrier network, from base stations via communications processors and into the server.
This potential collision of once-separate markets springs from the huge increase in processing tasks which infrastructure is required to handle. Base stations acquire applications processors to handle data as well as connectivity; and servers increasingly have to offload particularly intensive tasks from host processors to optimized coprocessors. This is where Texas Instruments sees its entry point to the cloud server world, the latest target for its Keystone II system-on-chip architecture.
The company insists it is not going all the way into the emerging market for ARM-based server host processors, as Marvell, Nvidia, Calxeda and others are, but it is still knocking on the door of the Intel fortress. It has already notched up a design win in a server, for a media coprocessor based on Keystone II, However, according to EETimes, Tom Flanagan ‘ director of technology strategy ‘ said: ‘There will be a lot of server-style apps we will be involved in, but directly replacing server blades is unlikely.’
Instead, servers will increasingly offload intensive tasks such as video transcoding and big data analytics from the main CPUs to coprocessors, an approach which can boost performance. Flanagan said that data centers were turning away from monolithic server designs in favor of ‘breaking job down and putting some things on a DSP or A15’.
It is also important to reduce power consumption, one of the reasons why ARM platforms (and Atom) are even breaking into the host processor in cloud servers. ‘Where you need heavy number crunching we’re very power efficient, delivering as much as 10 times the MIPS per power’ as x86, Flanagan claimed. Therefore, target markets will include high performance computers for applications like seismology and weather forecasting.
Keystone II harnesses up to four ARM Cortex A15 cores and multiple C66x DSPs, plus hardware acceleration for switching, packet processing and media. When the firm unveiled the upgraded architecture in February, it made it clear that, while the base station remained its primary target, it aimed to expand its platform into other types of infrastructure such as network processors and servers. The inclusion of packet processing and switching into the SoC makes it more suited to data-driven computing platforms.
Even before then, it had scaled KeyStone up to address the nascent demand for Cloud-RAN (as well as right down, into the embedded sector). The architecture was expanded for C-RAN and network servers, to target the convergence of cellular networks with massive data processing capacity. ‘Clustering base station functions allows processing capacity to be concentrated, which offers an innovative alternative to operators looking to lower operating expenses by reducing the amount of equipment in the field,’ said TI at the turn of the year.
Like Freescale, TI is looking to provide a single architecture and software base which run the whole RAN gamut from metrocell to macrocell, and then beyond the base station to the network server, which has traditionally been a separate segment. The specific enhancements which make KeyStone suitable for C-RAN include an expanded Multicore Navigator Queue Manager, which can provide over 16,000 queues; and an upgrade to the Hyperlink chip-to-chip interface to support 100Gbps with dual-port operation, while the integral wire-rate Ethernet switch is upgraded to 10Gbps per port. Flanagan said: ‘With KeyStone, we can now create device pools with unheard-of levels of capacity – nearly 800 cores pooled to appear as a single multicore device.’
Now TI has gone a step further, including new enhancements to stretch KeyStone further into the server. But Intel is also taking advantage of the rising use of coprocessors in all areas of infrastructure. It is surrounding its Xeon server CPU with a variety of different surrounding options optimized for different segments ‘ network processors initially, with base stations promised next year. It has also worked with small cell specialist Ubiquisys on a ‘smart cell’ which sits alongside the baseband and provides applications processing capabilities close to the user. And it is attacking the telco network from yet another angle, having acquired Fulcrum Microsystems earlier this year to fill another gap, in Ethernet switch-chips.
It is already a player in network processors and aims to expand the market for x86-based processors, particularly the Xeon server chips, to new sectors, as the traditional data center equipment space comes under pressure from reduced corporate budgets and ARM. Intel plans to turn out a series of companion chips for Xeon, each of them designed to replace the usual server-oriented coprocessors, with ones optimized for another markets, such as communications processors or even base stations. This will enable the giant to target x86 at applications previously addressed by specialized platforms.
One of the first instances was Jasper Forest, which paired a Xeon Nehalem processor with a modified server logic chip to target edge routers and storage, with considerable success. Together with a strong base in security appliances and the control plane, this has helped Intel become the second largest supplier of communications processors after Freescale, according to the Linley Group. It recently moved into the data plane, chasing Freescale, and next year plans a base station offering. This will partner Xeon with a companion chip targeted at digital signal processing in base stations, and 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).