Cloud computing is poised to shake up the enterprise industry in a big way, but could its impact reach all the way to the microprocessor?
A number of hopefuls certainly seem to think so, offering visions of new cloud computing platforms that will finally break the stranglehold that x86 architectures have on the modern data center.
One of these is Tilera Corp., which this week unveiled a new series of its Tile line that includes a 100-core device that the company is earmarking for everything from network and cloud systems to multimedia and wireless devices. The Tile-Gx family -- which can also be had in 16-, 36- and 64-core versions -- is built on a 40 nm process and offers 64-bit RISC cores, each with32k of L1 I- and D-cache, plus 26 MB L3-coherent cache. Each chip also has two or four DDR3 memory controllers, on-chip hardware acceleration and an integrated packet accelerator. The company claims a 10-fold improvement in compute efficiency compared with Intel's latest 32 nm Westmere architecture, even as it keeps power consumption to about 55 watts.
A couple of key design factors at work on the Gx line is the company's iMesh interconnect, a two-dimensional system that eliminates the need for an on-chip bus, and the Dynamic Distributed Cache (DDC) system that allows the total cache among all cores to be pooled. For cloud environments, this kind of flexibility allows the device to scale linearly according to the number of cores that are in place without encountering the data bottlenecks that usually accompany shared memory architectures. Tilera also provides a Multicore Development Environment, designed to help software makers overcome the threading difficulties that have hampered that industry's transition from single-core to multicore.
As we've seen in the past, that development community support makes or breaks any new hardware innovation. And since we're talking about enterprise applications, Tilera has taken on quite a task to get software companies like Microsoft to break out of the x86 mold. But that job might be easier than it seems now that Microsoft has identified manycore parallelism as one of the key enabling technologies for cloud-based data centers. The company has even gone so far as to set up the eXtreme Computing Group (XCG) to help foster the kinds of hardware, systems and platform development suitable for petascale computing environments. I'd be surprised if it isn't already taking a look at Tilera.
At the same time, however, Intel is ramping up production on its Clarkdale Nehalem processors, built on the Westmere architecture. Details on the design are still sketchy, but the few who have taken a look under the hood report that it's more than a shrunken version of the current 45 nm design. Among the new features hinted at by Chipworks is a new kind of NMOS strain, possibly using a carbon-doped source that would only be workable if Intel has figured out how to maintain its solubility in atomic-level lattices.
Whether Intel can deliver anything soon to outclass a 100-core design, even one that's built on a larger process, remains to be seen. But the company has the advantage of ingrained resistance among CIOs to shifting from x86 to RISC at this stage in the game, which may not last forever, but should provide enough breathing room to ramp up its manycore designs to production models.
No matter who takes the lead, however, it appears that data center processing capabilities are about to jump through the stratosphere.