GPU-based servers lack broad appeal
Graphics processing units (GPUs) have increasingly gained favor among companies that utilize high-performance computing (HPC), but programming challenges will limit the presence of such processors in many of today's servers, industry insiders noted.
According to Dennis Ang, business development director of HPC and service provider at Hewlett-Packard (HP) Asia-Pacific and Japan, not all business applications can run on GPU chips. Additionally, it will take time for existing apps to be redesigned and ported over to run on GPU-based servers. These two factors will hamper widespread acceptance of such servers, he noted in his e-mail.
An IBM spokesperson ZDNet Asia spoke to agreed with the HP executive's assessment. Sinisa Nikolic, high-performance computing program director at IBM Growth Markets said that "programming challenges" will continue to limit the adoption of GPU chips among server makers. While many high-end supercomputers have some graphics chip under the hood, these processers have not been widely taken up among enterprise server users, he noted.
However, with software advancement, this might change over time, Nikolic predicted in his e-mail.
On the current server landscape, Nikolic said that the vast majority of server deployments today are strictly central processing unit-based (CPU) without accelerators. CPU and GPU hybrids are a "distant second" while GPU-only servers are rarely seen, he added.
HP's Ang said that the company is working hard to change this composition. He said that HP has close working relationships with key GPU vendors and is involved in partnerships to integrate graphics-based accelerator technology into its server platforms.
For instance, HP in October announced its first three-GPU blade server. With the graphics chip in place, the server is adept at performing floating-point calculations, which makes it capable of improving graphics performance, cheap HPC functionalities and emerging business needs such as cloud computing and Web serving, according to an earlier report by ZDNet Asia's sister site ZDNet UK.
A good fit
Ang said that graphic processors are particularly gaining acceptance among organizations that need to accelerate their application delivery process. "GPUs are able to speed up delivery from between 10 to 100 times, depending on the application," he explained.
GPU chips are also useful in the high-performance computing arena, the HP executive pointed out.
This usefulness was highlighted when the China-based Tianhe-1A was named the top supercomputer in November. Liu Guangming, chief of the National Supercomputer Center in Tianjin, where the supercomputer is housed, was quoted as saying: "The performance and efficiency of Tianhe-1A was simply not possible without GPUs." In a separate report, a computer science professor said Tianhe-1A exploits the power of GPUs which are "awfully close to being uniquely suited to [the Linpack] benchmark [used to measure the TOP500 list of the world's most powerful supercomputers]".
Tapping onto the trend is cloud computing provider Amazon Web Services which added a GPU-based service for its Elastic Compute Cloud (EC2) in mid-November.