AMD Launches Project to Ease HPC Development on GPUs

The vendor's "Boltzmann Initiative" includes new compilers for C++ and CUDA, which will enable developers to port their software to AMD's platform.

Advanced Micro Devices initiative

Advanced Micro Devices is pushing a new initiative designed to expand the presence of its FirePro graphics technology in the high-performance computing space by accelerating application performance through software.

AMD is building on its work around heterogeneous computing to develop tools that officials say will increase the performance and efficiency of high-performance computing (HPC) systems that can leverage both CPUs and discrete GPUs. The offerings are part of what company officials are calling the Boltzmann Initiative—after Ludwig Boltzmann, who developed the theory of development of statistical mechanics.

The effort is the result of more than four years of work AMD has done in heterogeneous system architecture (HSA), in which CPUs and GPUs integrated on the same chip are treated as equals, and workloads that run on the chip can use either one based on its needs. It's the foundation of AMD's accelerated-processing unit (APUs), which offer the CPU and GPU integrated onto the same piece of silicon.

AMD, one of the founding members of the HSA Foundation, is looking to take what it's learned from that work and apply it to the HPC space and discrete GPUs, Greg Stoner, senior director of compute solutions technology at AMD, told eWEEK.

The chip maker unveiled the Boltzmann Initiative this week at the SC 15 supercomputing show in Austin, Texas.

GPUs have become an increasingly important tool in the HPC space for driving system performance and energy efficiency. Organizations for almost a decade have used GPUs as accelerators that can run parallel code and free up the CPU for other applications. AMD and Nvidia have been the primary suppliers of the GPU accelerators. Nvidia, with its CUDA environment, has been the more dominant player in the space, besting AMD and OpenCL, which also is supported by such tech vendors as Intel, ARM and Qualcomm.

According to the latest Top500 list of the world's fastest supercomputers, 104 of the systems use an accelerator technology, an increase over the 90 on the list in June. Of those, 63 use Nvida GPUs, and 27 use Intel's x86-based Xeon Phi coprocessors. Three use AMD's Radeon GPUs, and four use a combination of Nvidia GPUs and Xeon Phis.

AMD is looking to change the picture through its efforts in software. A key will be a new HSA extension that can bring similar capabilities to AMD's discrete FirePro GPUs as are found in integrated chips, including enabling the GPU to address the GPU and CPU memory as a single resource pool. It will bring greater support for Linux in the form of a new 64-bit driver for headless Linux. The extension—HSA+—won't be adopted by the HSA Foundation as part of its standard, though AMD may offer it to the open-source community later.

The chip maker also is introducing the Heterogeneous Compute Compiler (HCC), which will be able to compile for both the CPU and GPU, rather than requiring separation between the two. The C++ compiler will give developers a simpler environment with single-source execution, enable them to better create code for parallel execution and automate the placement of code to whatever processing elements is best suited to its needs.

In addition, AMD is introducing the Heterogeneous-compute Interface for Portability (HIP), a tool that essentially creates a bridge that allows for applications written for CUDA to be translated to common C++, eventually enabling those applications to run atop AMD graphics technology. At the supercomputing show, AMD is demonstrating a CUDE-generated Rodinia benchmark suite on AMD GPUs. According to AMD officials, about 90 percent of CUDA code can be automatically converged to C++ by HIP, and the other 10 percent can be converted manually. Software that could only run exclusively on CUDA can now be converted to C++ and run on AMD's GPUs.

In the same sense, software developed for AMD can be ported to CUDA.

"It's a new pathway for [CUDA] programmers to get onto our platform and use our tools," Stoner said.

As AMD officials looked at the HPC space, they found it was better for the company to expand its software portfolio and to broaden what developers could do with its software tools.

"We realized we needed to bring choice," AMD Fellow Ben Sander told eWEEK. "Choice is crucial."

AMD's announcements don't mean the chip maker is moving away from OpenCL, officials said. Also at SC 15, the company announced it has expanded its GPU compute libraries with the addition of HcBLAS and HcFFT.

The chip vendor is planning an early-access program for the Boltzmann Initiative tools for the first quarter of 2016.