High-performance accelerators for HPC environments – such as AMD FireStream or Nvidia Tesla – can eventually find themselves inside mainstream servers, but it will not happen shortly.
Graphics processing units (GPUs) are making inroads into high-performance computing (HPC) markets these days and their performance promises rather bright future for general purpose processing on GPUs (GPGPU) in general. But if their horsepower is that high and the benefits are visible, why not install them into mainstream servers? Not until the appropriate software is ready, claims AMD; in fact adoption of GPGPU technologies by mainstream cloud or data-center servers may take longer than the adoption of 64-bit microprocessors.
“The most initial interest right now from the server perspective has been on HPC. Those people have specialized code that can leverage a specific way that GPUs are designed to provide compute capabilities. There is tremendous excitement about stream computing in the high-performance computing segment. As it traditionally happens, HPC adopts a technology first and takes advantage of it and then it will slowly make its way into the mainstream,” said Brent Kirby, senior product manager at AMD.
The most important thing for mass adoption of GPU stream processing is software that can take advantage of graphics chips. There are not a lot of such programs even for client computers and there are virtually none non-proprietary solutions for servers. As a result, companies like AMD or Nvidia Corp. have to work hard on creation of tools that can leverage the power of GPUs in non-graphics applications and then encourage developers to design GPGPU-based software.
“When it comes to GPGPU, a lot depends on software, which is why we are really focused on OpenCL and Open64. Getting the software ecosystem in place so people can develop code that can take advantage of all these different compute units that might be together someday is crucial,” added Mr. Kirby.
In fact, pushing GPGPU into the mainstream server space will not be the first time for AMD to promote something brand-new there. When AMD released its x86-64 Opteron and Athlon 64 processors back in 2003 there merely were operating systems along with a handful of programs to support 64-bit chips. Eventually, the transition to x86-64 from 32-bit happened pretty quickly on the mainstream server market. Only time will tell whether stream computing will also take off quickly.
“When we launched the 64-bit Opteron processor for servers, we had some Linux [versions] and we also coincided with Windows Server 2003 64-bit OS. It took a while for that ecosystem to get put in place, but it happened pretty rapidly. People could take instant advantage of 64-bit systems: memory addressability capabilities mean a lot for servers. I do not think that stream computing will take off as quickly as 64-bit technology, but I think it is going to start off in HPC and work its way to more traditional workloads,” noted the senior product manager at AMD.
In fact, there is interest towards stream processing from cloud computing segment, but it does not come from a lot of customers.
“I think GPGPU will be a part of cloud/hyper-scale datacenters, but it will not be in the majority of them initially. I know there is interest from the cloud computing with regards to CPU and GPU together, but it does not come from a large portion of end-customers yet,” said Brent Kirby.