Looking beyond iPhone, ARM 64-bit chips get into servers with GPUs
- 23 June, 2014 17:04
Breaking from the cocoon of the iPhone 5S, 64-bit ARM processors will start delivering breakthrough performance in servers, aided by graphics cards used in some of the world's fastest computers.
The first high-performance servers with ARM 64-bit processors have been announced with Nvidia's Tesla graphics cards, which is also in the U.S. Department of Energy's Titan, the world's second fastest supercomputer. The servers from Cirrascale, E4 Computer and Eurotech were announced at the International Supercomputing Conference in Leipzig, Germany.
The first 64-bit ARM processor was used by Apple in the iPhone 5S, which was introduced last year. No 64-bit ARM products have been announced since, but there is a growing interest in low-power ARM servers to process lightweight tasks such as responding to search and social networking requests.
Hewlett-Packard, Dell and others have plans to ship 64-bit ARM servers that could help cut electric bills in data centers. ARM processors alone cannot deliver the horsepower necessary for complex scientific and math calculations, while GPUs can speed up such tasks.
There is an interest in ARM processors combined with GPUs in research areas like protein folding, drug discovery and atomic simulations, said Ian Buck, vice president of accelerated computing at Nvidia.
"They now have an alternative than x86 servers, they can validate more choice," Buck said.
Most supercomputers today use processors from Intel or Advanced Micro Devices based on the x86 microarchitecture, which is also in PCs. But based on historic trends, researchers have argued that smartphone chips could ultimately replace the more expensive and power-hungry x86 server processors in supercomputers.
The first ARM servers from Cirrascale and E4 Engineering will ship later this year. The Cirrascale RM1905D and the E4 EK003 have eight-core AppliedMicro X-Gene processors and Tesla K20 GPUs, with support for up to DDR3 memory, 10-gigabit Ethernet and PCI-Express 3.0. The systems come with 400 watt power supplies. The Cirrascale 1U server is for cloud and high-performance applications, while the E4 Engineering 3U server is for Web computing, analytics, video rendering and science applications. Configuration details for Eurotech's ARM server was not available, but it will have liquid cooling, according to Nvidia.
For Nvidia, ARM servers represent a big opportunity to sell its graphics cards. The Tesla GPUs are already compatible with x86 chips and IBM's Power processors, and support for ARM-based chips is the next logical step, Buck said.
"There are lot of interest in ARM64," Buck said.
AppliedMicro's X-Gene chip is based on its proprietary chip design, and has server features such as error correction and RAS (reliability, availability and serviceability), which are typically not available in ARM mobile chips. The X-Gene also has I/O, networking and signal-processing components.
But ARM servers lack software support as most server applications are written for x86 chips. But Hadoop, OpenStack and the LAMP stack already support ARM, and native Java support is coming in 2015.
Nvidia at ISC is also announcing CUDA 6.5, a set of proprietary parallel programming tools that can harness the joint computing power of CPUs and GPUs. CUDA 6.5 adds support for ARM processors.
Beyond AppliedMicro, Nvidia is keeping a close watch on other ARM-based server chip makers such as Broadcom and Cavium, Buck said. He hinted that GPU support could also come for those chips.
Nvidia faces competition from AMD, which is developing an ARM server processor and sells graphics processors such as FirePro for the supercomputing space. Nvidia lacks a CPU specifically for servers and is not developing one, Buck said.