AMD takes the next step in heterogenous computing
SAN JOSE, Calif. — The Heterogeneous Systems Architecture (HSA) Foundation is making good progress in an area important for future designers of SoCs and the software that rides on them, but it still faces a long road to success.
This week, Advanced Micro Devices said it will ship the first processor to use HSA techniques of letting CPU and graphics cores share resources in January. AMD developed support under Linux for the techniques, but no other plans have been announced so far for processors or operating systems supporting the techniques. The chips (and probably the operating systems) will come at some point.
ARM, Imagination, Mediatek, Qualcomm, Samsung, and other top mobile SoC and core designers are part of HSA. During a panel discussion at AMD's developer conference here, an Imagination executive said the company will support HSA on all its cores eventually, but he would not give say which cores would ship when. ARM CTO Mike Muller was more reserved. "HSA is good for midrange systems, but I'm not sure if it scales to the smallest systems, and I don't think it does scale to high-performance computing with tens of thousands of processors," he said during the panel discussion. "One system's architecture won't scale across the whole space."
Apple, Google, and Microsoft have remained mum about their plans to support HSA in iOS, Android, Chrome OS, and Windows. Designers can write drivers to support the techniques in operating systems such as Windows, but that's not an ideal approach.
The big standouts are AMD's archrivals: Intel and Nvidia. They and Apple back OpenCL in mobile and desktop systems. The Khronos Group's OpenCL spec is a higher-level API for programming parallel tasks on any GPU. It has been around for a while but is just starting to build in some of the shared-memory goodies the HSA Foundation is defining.
A representative for Khronos and Nvidia said HSA, OpenCL, and Nvidia's proprietary Cuda software environment are on parallel paths and offering increasingly similar features.
Cuda has been a hit for Nvidia in high-performance computing, including in the top 500 supercomputers. Thanks to an early move into this market, a growing share of these systems use Nvidia's GPUs with Cuda. But that could change.
"Every three or four years, supercomputer developers look around for what's new, seize on something, and really milk it for a few years," said Wen-mei Hwu, a professor of computer science at the University of Illinois and an expert in the field. The latest versions of Cuda already support some HSA flat-memory features. But Cuda is not well suited for some supercomputer apps, and that could open a door for the HSA Foundation and AMD.
To read more, go to “Intel’s Phi and AMD’s GPUs.”