Nvidia buys Portland Group for compiler smarts
C++ and Fortran to span ARM and GPU ceepie geepies
Graphics chip maker Nvidia has big aspirations to get into computing proper with ARM processors and GPU coprocessors, and its odds in its battle against archrival Intel may have just gotten a lot better now that it has snapped up The Portland Group.
The financial terms of the acquisition, which has been completed, were not disclosed.
PGI, as the company is known, was founded in 1989 and kicked out Fortran and C compilers for Intel's i860 RISC processors two years later. It has been a driving force behind the development of parallel Fortran compilers over the years.
It was tapped by Intel to do the Fortran for the ASCI Red massively parallel supercomputer at Sandia National Laboratories in 1996 and the first machine to break the teraflops performance barrier.
PGI also did the compilers for the "Red Storm" machine built by Cray using Opteron processors from Advanced Micro Devices and the "SeaStar" interconnect developed by Cray to lash them together.
The company has been very good at seeing and riding changes in processor or coprocessor technology and was out in front with support for OpenMP for Linux, SSE/SIMD engines in x86 processors, 64-bit x86 processors.
In recent years, it has been the Fortran supplier for Nvidia's CUDA GPU programming environment and has created a set of compilers that allow for CUDA to dump code onto multicore and multithreaded x86 processors.
Significantly, PGI was one of the partners that joined Nvidia in setting up OpenACC, which is trying to establish an open standard for adding directive hints to compilers to help them parallelize applications for CPUs, GPUs, and any other kind of parallel execution engine (such as a Xeon Phi from Intel). The company also last year launched an OpenCL compiler for multicore ARM processors.
With Nvidia working on its "Project Denver" ARM processor, those ARM skills are going to come in handy. And rather than just partnering tightly with PGI, as Nvidia has been doing, the company has decided that it needs to control a compiler stack.
This makes sense. IBM has always controlled the compilers on its proprietary and Power processors, and Intel followed suit and has control of its own compilers, too.
Both companies are happy to have others make compilers for their chips, of course, but the important thing is to have a set of your own compilers that can be tweaked alongside chips as they change. In a world where performance is everything, the compiler is often the deciding factor. (And sometimes, captive compiler makers put a thumb on the scale and they usually get caught, too.)
Nvidia has a software development team that spans 2,400 software engineers, and some of them work on compilers and other aspects of application development. But Buck says it is hard to extract who is working explicitly on GPU computing because the Nvidia software team is "highly leveraged" across all aspects of GPUs.
"What we don't have is a world-class HPC compiler team on the scale and with the kind of products that PGI is offering,' Ian Buck, general manager for the CUDA compiler stack at Nvidia, tells El Reg. "By working as one company, we can now better align our technology roadmaps and hopefully accelerate our innovation around GPU computing,"
He also stressed that Nvidia would continue to work with TotalView, CAPS, Cray, Allinea, and other compiler partners, and that nothing would change in this regard in the aftermath of the PGI acquisition.
And even more importantly, with Nvidia doing its own server-class ARM processors and aiming them at high performance computing jobs, it is going to need good compilers for that work on ARM chips and can offload work to GPU coprocessors.
"PGI has experience with ARM," says Buck, "but there is no commercial Fortran compiler available – yet."
He was quick to add that Nvidia was not pre-announcing any product plans, of course.
The PGI team will remain in Portland and that all 30 people working on various compilers and development tools will start getting their paychecks from Nvidia.
PGI will remain a wholly owned subsidiary of Nvidia, as it has been a subsidiary of chip maker STMicroelectronics since 2000. PGI has about 5,000 sites worldwide using its compilers and development tools. ®