SAN JOSE,Calif. — The Heterogeneous SystemsArchitecture (HSA) Foundation ismaking good progress in an area importantfor future designers of SoCs and thesoftware that rides on them, but it stillfaces a long road to success.
This week, Advanced Micro Devices saidit will ship the first processor to use HSAtechniques of letting CPU and graphics coresshare resources in January. AMD developedsupport under Linux for the techniques, butno other plans have been announced so farfor processors or operating systemssupporting the techniques. The chips (andprobably the operating systems) will come atsome point.
ARM, Imagination, Mediatek, Qualcomm,Samsung, and other top mobile SoC and coredesigners are part of HSA. During a paneldiscussion at AMD's developer conferencehere, an Imagination executive said thecompany will support HSA on all its coreseventually, but he would not give say whichcores would ship when. ARM CTO Mike Mullerwas more reserved. “HSA is good for midrangesystems, but I'm not sure if it scales tothe smallest systems, and I don't think itdoes scale to high-performance computingwith tens of thousands of processors,” hesaid during the panel discussion. “Onesystem's architecture won't scale across thewhole space.”
Apple, Google, and Microsoft have remainedmum about their plans to support HSA in iOS,Android, Chrome OS, and Windows. Designerscan write drivers to support the techniquesin operating systems such as Windows, butthat's not an ideal approach.
The big standouts are AMD's archrivals:Intel and Nvidia. They and Apple back OpenCLin mobile and desktop systems. The KhronosGroup's OpenCL spec is a higher-level APIfor programming parallel tasks on any GPU.It has been around for a while but is juststarting to build in some of theshared-memory goodies the HSA Foundation isdefining.
A representative for Khronos and Nvidiasaid HSA, OpenCL, and Nvidia's proprietaryCuda software environment are on parallelpaths and offering increasingly similarfeatures.
Cuda has been a hit for Nvidia inhigh-performance computing, including in thetop 500 supercomputers. Thanks to an earlymove into this market, a growing share ofthese systems use Nvidia's GPUs with Cuda.But that could change.
“Every three or four years, supercomputerdevelopers look around for what's new, seizeon something, and really milk it for a fewyears,” said Wen-mei Hwu, a professor ofcomputer science at the University ofIllinois and an expert in the field. Thelatest versions of Cuda already support someHSA flat-memory features. But Cuda is notwell suited for some supercomputer apps, andthat could open a door for the HSAFoundation and AMD.
To read more, go to “Intel’s Phi and AMD’s GPUs.”