Intel, ARM Take Competition Into HPC Arena

By Jeffrey Burt  |  Posted 2014-06-23 Print this article Print
servers data center

Meanwhile, Nvidia officials announced that its Tesla K20 GPU accelerators with the vendor's CUDA 6.5 parallel-processing platform support X-Gene in systems aimed at the HPC space, and that the first GPU-accelerated ARM64 development platforms will be rolling out from system makers like E4 and Cirrascale in July.

Officials with both Applied Micro and Nvidia said 64-bit ARM SoCs are a good fit for the HPC space, where power efficiency is increasingly important and recompiling software for particular architectures is nothing new.

"There is definitely a need for power efficiency, so HPC guys are incentivized" to explore alternatives to x86 chips, Singh told eWEEK. Roy Kim, marketing manager for Nvidia's Tesla Group, agreed.

"HPC is definitely the most flexible out of all data center customers," Kim told eWEEK. "They're used to recompiling [their applications] all the time."

However, with Intel, ARM and its partners are competing against a highly innovated, well-funded and determined company. Intel has leveraged its Atom platform to contend with ARM's expected encroachment in the dense, low-power server space. Now the company is making significant strides with its next-generation Xeon Phi chips.

The new Knights Landing chips will include the company's Silvermont microarchitecture, which currently is found in Atom SoCs and will help drive up the performance of the processors while reducing the power consumption. It also includes new on-package memory being developed with Micron Technology that will offer five times the bandwidth of DDR4 in a third of the space and with five times the power efficiency. It will start off with 16GB in the new memory.

At the same time, Knights Landing will feature a new fabric technology that Intel officials call Omni Scale that also will be integrated with the 14nm generation after Knights Landing. Rajeeb Hazra, vice president of Intel's Data Center Group and general manager of its Technical Computing Group, would not go into many details about Omni Scale, saying only that it wasn't Infiniband—which Intel is using today—but is compatible with it. Creating an interconnect fabric has been a focus of Intel for several years, and a driving force behind such acquisitions as technologies from Cray and QLogic, as well as Fulcrum Microsystems.

Intel's Wuischpard also said that the Omni Scale fabric technology will be open, enabling other vendors to build off of it.

The Intel officials said Knights Landing will have more compute cores than the 60-plus in the current Knights Corner, but would not say exactly how many. However, with all the enhancements and new technologies, they said it will offer more than 3 teraflops of performance in a single package, more than the 1.2TF in the current generation of Xeon Phi.

The key is moving beyond just offering a compute core by integrating everything from the interconnect and power efficiency capabilities to the memory and storage onto a single chips, Hazra said.

"It's not just the processor performance, but all these other parts as well," he said.


Submit a Comment

Loading Comments...
Manage your Newsletters: Login   Register My Newsletters

Rocket Fuel