Moments ago NVIDIA unleashed news that it would be acquiring the Portland Group International (PGI), which has been a standby in the HPC market for Fortran and C compiler tools.
PGI spun out of IP acquired from Floating Point Systems in 1989, and managed to carve out an early niche in Fortran and C compilers and tooling. From the early days pushing to the Intel i860 crowd up to the modern core-heavy chips it supports for its HPC-oriented base, PGI has expanded well beyond basic architectures and has become a particularly close partner with NVIDIA in recent years.
The Portland Group has maintained special focus on the needs of HPC markets, including its instrumental role in bringing high performance Fortran (HPF) to bear, which is an extension that opened trusty Fortran to a wider set of architectures. They have also been prominent backers of GPGPU technology from its early days. From the development of their CUDA Fortran in conjunction with NVIDIA and now their push to bring OpenACC to a wider community, it’s clear how the GPU company saw them as a prime fit.
According NVIDIA’s GM of the Tesla unit, Sumit Gupta, PGI has been an integral part of its strategy in GPU computing. Specifically, this is clear via its CUDA Fortran (which NVIDIA used in favor of rolling its own) and more recently, its ongoing support for OpenACC. Interestingly, during our brief conversation in advance of the news, Gupta stressed OpenACC as the shiny part of the deal, noting that while it has seen significant adoption, more work and emphasis in that area will result in expanded GPU adoption.
NVIDIA has been investing in its software ecosystem significantly, Gupta says, noting that without the developer community on board their technology can’t progress. NVIDIA sees OpenACC as the panacea (at least against the Xeon Phi quick-port buzz) since it makes it easier to tap into GPU acceleration.
He told us that with a higher-level approach to taking GPUs for a spin, it’s now typically a matter of 2 week to 2.5x speedups (on average) or the fruits of a one-day hands-on workshop with actual code to hit those same results. The problem with OpenACC isn’t functionality, he says, it’s simply around messaging--getting the word out that OpenACC offers a faster, cleaner way to using a higher-level approach to getting results out of GPUs, which have suffered a bad rap in terms of on-boarding.
We asked Gupta, about what competitive advantage this acquisition lends them over Intel, which has its own suite of x86 tooling and compilers. He downplayed the role of competition in this acquisition, saying that this was more of a long-term investment in the software ecosystem for their GPU. Although it’s hard to ignore a future where the all-important software angle (especially once we look toward a more diversified future processor-wise) is not “competitive” at this point, NVIDIA is still building its base to firm up how it delivers its GPU message.
As Michael Feldman, Senior Analyst with Intersect360 Research commented, "As accelerators hit the mainstream, the real challenge will be to make these devices comfortably programmable for the average coder. To do so, I think NVIDIA has realized that they'll need compiler technology beyond their CUDA toolset. This is what PGI gives them: C and Fortran compilers with a high-level programming framework geared to accelerators. And, like their main rival Intel, NVIDIA wants to this technology in-house."
Feldman looked at the competitive ecosystem as well, pointing to the ways that PGI assets can help NVIDIA with its ARM-GPU strategy via Project Denver, which could shake up the workstation and server market beginning late next year. In this case, as with other server business moves, it’s less about the metal than the mush.
On that note, all chipmakers are being taken to task when it comes to their software stacks and tooling, both at the general datacenter level and for their HPC user bases. Last week we listened as Intel spelled out its plans to continue pushing new capabilities into their Xeon line and beyond and now, others, including NVIDIA, are taking taking aim at the same target.
Despite the emphasis on OpenACC with this acquisition, NVIDIA says that continued maintenance and updates to CUDA Fortran are key to their strategy. Also, rather interestingly, they are keeping nearly everything status quo at PGI, at least for the time being. And of particular note there is the fact that they are going to continue supporting their x86 line of compilers and tools--the very tooling that serves as a competitive threat.
Gupta doesn’t see this as a problem, stressing once again the role that both the CUDA Fortran and OpenACC angles mean for their future growth. In other words, the conversation was about future directives versus any short-term benefit. And they’re taking OpenACC very, very seriously.
PGI will continue to operate independently as a unit inside NVIDIA and its staff reporting to Ian Buck, GM of GPU Computing at NVIDIA. One can imagine a day when the x86 products might dwindle, but that’s just speculation that’s being bandied about.