Multiprocessing isn’t new. Two or tens of processors have been hitched up in research facilities and server farms for decades. But replicating cores on a single piece of silicon is a fairly new phenomenon. These multicore processors aren’t just the odd anomaly, and not a mere trend, but in late 2010 multicore has turned into a movement – a full-fledged 1960’s Arlo Guthrie society-changing movement.
The last year has seen essentially every major processor company add or expand multicore processors in their repertoire. A few are improved versions of earlier dual cores while others are fresh new multiple-births. Why are there so many multicore offerings now? The simple answer is that to stay on the performance treadmill that Moore’s Law so aptly described, multicore is the only viable option. But there are different approaches within the multicore movement, and many have been utilized for years for a more delicate operation, perhaps right in your pocket.
Multicore is upon us
Any system designer that thinks multicore is out of his realm just isn’t paying enough attention. Like peace, liberation, civil rights, and (yes) saving the Earth in the ’60’s, almost imperceptibly, ideas took shape, action was taken in small pockets, and before you knew it, widespread change – a real movement – had actually taken hold and was showing results. Look up and you’ll notice multicore processors are all around.
Indeed, a fevered pitch is coming from chip and intellectual property (IP) vendors alike as they scramble to proclaim more, newer, better, and faster multicores than the other guy. The traditional modest performers that focus first on battery life are utilizing multiple cores to reach performance levels high enough to be real threats to entry markets of the powerhouses. The typical “big iron” frequency chasers are using multiple cores to reign in the heat they’ve been generating to keep the low-power challengers at bay yet be able to keep the server farms pumping out ever-more search results, Web pages, and video feeds.
Just Monday, MIPS Technologies revealed a new multicore IP offering that extends the 1004K Coherent multi-processing family to what is now called the 1074K Coherent Processing System. These old hands at multiprocessing are offering up to four MIPS32 74K cores in a coherent synthesizable platform that can run 15,000 Coremarks at 1.5 GHz on a 40 nm process and probably up to 2.5 GHz when carefully crafted at 28 nm. MIPS believes that just three of their cores fit in the same die area of an equivalent Intel Atom core yet can run 2.5x faster. The 1074K is essentially available for licensing now so a vendor or OEM can start building a chip around one immediately. MIPS expects the 1074K to invigorate its position in set-top box (STB), digital TV, Blu-ray players and home/office networking products, enhancing the linkage to networks and the Internet.
The industry has its own Octo-Mom now with Intel flashing fuzzy images earlier this month of an 8-core outgrowth of its current quad-core Xeon processors that will have a birth date in 2011. Such behemoths and three-headed monsters from AMD are designed for server farms, high-performance computing (HPC) applications, and maybe some networking applications that can consume all the performance they can get.
At the other end of the spectrum, tallying its success with “Intel Architecture” in the low power arena, Intel expects that 2 million of its Atom processors will be shipped in the embedded markets in the three years since its introduction, plus another 70 million in Netbook/MID-type markets. Intel also has a dual-core version of Atom to shore up its strength against an Army of competitors.
Early in September, ARM Ltd with Texas Instruments by its sidegave a number to “Eagle”, their formerly-secret high-end Multicoredesign. The Cortex-A15 will be another year in development (as IP) butpromises untold performance coming from a pair of greatly enhanced ARMcores. Like the most high-performance cores it will make use ofout-of-order superscalar processing and carefully managed Level-2caches. Four-core configurations and a new 40-bit addressing capabilityadd street cred to ARM as a viable consideration to traditional higherperformance computing systems.
Besides cementing the dominance of the ARM architecture inanything resembling a cellular handset, the A15 is the most recentweapon to hammer away at the power-wise server and wirelessinfrastructure applications. The Multicore Cortex-A9 has just begungetting its pair of cores into chips, so the possibly-2.5 GHz A15 willbe a nice follow-up. What TI which has OMAP put ARM on the map could dowith A15 fascinates the imagination.
Applications Processors, Heterogeneous
A flurry of those dual-core ARM Cortex-A9 designs wereannounced just before summer. Samsung previewed a gigahertz A9 designto upgrade the Hummingbird applications processor destined for moreadvanced cellular handsets. This “Orion” should hit 1 GHz thanks to a45 nm process node which should boost performance for such non-telephoneuses as 3D graphics and HD 1080 video.
In May, Qualcomm also revealed an ARM-based applicationsprocessor (it has an architecture license) to upgrade its Snapdragonwith two heads. Running 1.2 GHz and coming from Qualcomm, thesechipsets have special support for UMTS and CDMA in a cellular baseband. A slightly more advanced design is expected later this year.
Marvell just announced a “tri-core” update to their ARMADAapplications processor series. It’s really a 1.5 GHz dual-core ARMprocessor (Marvell has an ARM architectural license) and is supplementedby a 624 MHz presumably-ARM core that can handle special low-powerduties. Like most applications processors this has a number of otherspecial processors and accelerators to greatly improve efficiency. Thisone layers in engines for 3D graphics, two-stream 1080p video, audio,advanced cryptography, and image signal processing. The ARMADA 628 isstarting to sample.
Even STMicroelectronics has created dual-core ARM applicationsprocessors, initially using the old ARM926 (at just 333 MHz) but isworking on a Cortex-A9 dual-core version of their SPEAr that will runtwice as fast.
In the end, these ARM-based applications processors requiredetailed lists to compare features and capabilities, but their dexterityis impressive. Freescale has been quiet on doubling up the ARM core onits i.MX processors (at the heart of the Kindle e-book), but most ofthese common heterogeneous multicore SoCs are going a little bitsymmetric multiprocessing too to get the overall general processingheadroom needed to jump into the less-clear future and to expand tobroader markets. The trend is clear.
Some applications can take advantage of dozens of smaller,more-dedicated processors breaking down predetermined tasks en masse. These very fast cores often come in large enough quantities to bereferred to as Many-core rather than just multicore. A few MIPS coreswere used in this way in years past, as were some high-count ARC andTensilica core applications which similarly took onsufficiently-consistent applications with the efficiency of antsdevouring fallen trees. Some instruction set architectures (ISA) aredesigned specifically for this approach, such as the 64-core matricesoffered by Tilera, with promises of 100-core chips in 2011. Last JuneTilera declared that every year it would double the number of cores inits chips, with end-uses like high-performance computing (HPC), cloudcomputing, data centers, and data communications.
It’s a movement
One could say the evidence is overwhelming that multicore isupon us. It is coming in many varieties and to serve many purposes. The last few months have seen a groundswell of activity aroundmulticore. If the year was 1967 again, the thoughtful people in thestreets might declare Multicore to be a movement, pick up a guitar, andwrite a song about it, or organize a protest. By now it should be soobvious that even the engineers should understand. Do not let multicorepass you by. Embrace it.
About the author
Tom Starnes is processor analyst at Objective Analysis. He will be moderating EE Times Virtual Conference entitled “Approaching Multicore,” Thursday, September 30, 2010.