AMD Speeds Up OpenCL on APUs with Latest SDK

AMD has just released an updated version of their APP SDK, bringing it up to version 2.5. This latest release features a number of performance enhancements designed to exploit the increased CPU to GPU bandwidth available in the company’s APUs, in some cases achieving effective data transfer rates as high as 15GB per second with the latest AMD A-Series APUs, per the press release.

“Improving performance and programmability on AMD platforms is a key initiative for us, and we work closely with developers to help us make the APU the best possible development platform,” said Manju Hegde, corporate vice president, AMD Fusion Experience Program. “With these latest SDK refinements, we are supporting continued growth of the ecosystem and building on the momentum generated by the successful AMD Fusion Developer Summit held in June of this year.”

Part of the performance improvements are made possible due to the fact that the CPU and GPU segments in an AMD A-Series APU share a common DDR3 memory pool, with the memory partitioned between the CPU and the GPU. Unlike a discrete GPU which must transfer data over PCI Express to the CPU, with an APU that data may be transferred more quickly between the CPU and the GPU as there is no longer a constraint imposed by PCIe.

With the AMD APP SDK 2.5, the company has introduced a zero copy transfer path for buffers defined by using standard OpenCL guidelines. In addition, the AMD APP SDK provides multi-GPU support on Windows platforms including support for APU plus discrete GPU configurations that enables compute performance scaling across multiple GPUs, as well as AMD PowerXpress technology support for APU plus discrete GPU configurations.

One of the common threads of the AMD Fusion Developer Summit was obviously heterogeneous computing and more easily developing applications that can seamlessly leverage the power of both the CPU and GPU resources available in a system. This latest AMD App SDK release takes the company a step closer to achieving some of the goals set forth at the summit.

Marco Chiappetta

Marco Chiappetta

Marco's interest in computing and technology dates all the way back to his early childhood. Even before being exposed to the Commodore P.E.T. and later the Commodore 64 in the early ‘80s, he was interested in electricity and electronics, and he still has the modded AFX cars and shop-worn soldering irons to prove it. Once he got his hands on his own Commodore 64, however, computing became Marco's passion. Throughout his academic and professional lives, Marco has worked with virtually every major platform from the TRS-80 and Amiga, to today's high end, multi-core servers. Over the years, he has worked in many fields related to technology and computing, including system design, assembly and sales, professional quality assurance testing, and technical writing. In addition to being the Managing Editor here at HotHardware for close to 15 years, Marco is also a freelance writer whose work has been published in a number of PC and technology related print publications and he is a regular fixture on HotHardware’s own Two and a Half Geeks webcast. - Contact: marco(at)hothardware(dot)com