AMD Aims for GPUs in Mainstream Servers Starting 2012
Advanced Micro Devices will put more focus on tightly integrating graphics processor cores into mainstream servers starting 2012 as it tries to increase system performance, a company executive said.
Mainstream servers in the future could have a combination of graphics processors and CPUs in servers as applications take advantage of thousands of GPU cores, said Gina Longoria, director of the product management and workstation division at AMD. The company may provide CPUs and GPUs together in a server to run highly parallel applications, she said.
Graphics cards are considered more advanced in tackling scientific and math applications than CPUs, which are designed to handle generic computing tasks. Some servers already combine the performance of GPUs with CPUs in high-performance computing environments.
For example, supercomputing vendor Appro offers the HyperPower Cluster, which is designed to break up and simultaneously execute thousands of threads and tasks across servers with separate Intel CPU cores and Nvidia GPU cores. One Nvidia GPU-based supercomputer, called Tsubame, at the Tokyo Institute of Technology has more than 30,000 processing cores that provide a processing speed of 77.48 teraflops.
But while the CPU and GPU are combined mostly on supercomputers, many PCs also come with basic parallel processing functionality to boost application performance. Laptops today can unload specific multimedia tasks like video editing directly to graphics chips, leaving CPUs free to execute tasks like word processing and virus scanning. AMD hopes to bring such mixed computing environments to servers in the future, Longoria said.
Over the next two years, computing using a GPU will remain a relatively niche area, so AMD's focus will be on CPUs and adding cores to it, Longoria said. Over time there could be tighter integration, and AMD could de-emphasize CPU cores if heterogeneous computing using GPUs takes off, Longoria said. AMD currently has a 16-core server CPU code-named Interlagos planned for release in 2011, and it will be based on a new chip architecture.
"As GPU becomes more relevant, that's a better way of getting performance than [CPU] cores," Longoria said. However, there's no hard and fast rule because GPU-based computing may not always be the right tool unless applications are written in parallel, she said.
Graphics-intensive applications that involve face recognition, video and imaging will see improvement on servers with the help of GPUs, Longoria said. She couldn't comment on whether data-intensive applications like databases would see improvement with the help of GPUs.
GPU is the fastest way for servers performance to jump, said Dan Olds , principal analyst at Gabriel Consulting Group. The combined CPU-GPU use is growing in high-performance computing environments, and will soon come to businesses that do a lot of graphics intensive and analytical tasks, he said.
"I'm glad they are addressing the market, but perhaps they should push ahead and develop the market more," Olds said. AMD is more of a spectator and not a major player like Nvidia, which is aggressively pushing its software and hardware for heterogenous computing, he said.
Nvidia is set to release its newest graphics processor based on the Fermi architecture with 512 cores.
AMD is taking many steps in order to harness the parallel processing capabilities of CPUs and GPUs to accelerate application performance, Longoria said. The company is providing tools to encourage development of applications using the OpenCL standard, which includes a C-like programming language with APIs (application programming interfaces) that enable parallel task execution across hardware including CPUs and GPUs. Apple, Intel and Nvidia Devices are among the companies promoting OpenCL.
A lot of the high-performance computing community is doing their work based on the OpenCL platform, which could eventually trickle down into general-purpose computing, Longoria said. OpenCL competes with Microsoft, which is promoting its proprietary DirectX parallel programming tools, and Nvidia, which offers the CUDA framework.
AMD is focusing OpenCL platform applications to be written for the FireStream processor, a family of high-performance GPUs. AMD acquired graphics technology when it bought ATI Technologies in 2006.
AMD designs both CPUs and GPUs, giving it a leg up over the world's leading chipmaker Intel, whose CPUs go into close to 80 percent of PCs worldwide. Intel in December unceremoniously delayed its first graphics processor called Larrabee without providing a new release date, leading to speculation that the product had been scrapped.