Nvidia’s CEO, Jen-Hsun Huang, has slammed Intel’s unreleased OneAPI software programming approach. During the company’s Q2 2019 earning’s call, Huang explained that he doesn’t “really know how one programming approach or simple API is going to make seven different type of weird things work together… programming isn’t as simple as a PowerPoint slide.”
Nvidia just wrapped up its Q2 earnings call, which were largely positive for the green team and its investors. During the call, Huang was asked by Cowen’s Matt Ramsey whether he saw any movement in the competition when it came to datacentre software stacks. And Huang had quite a bit to say on Intel’s unified programming model.
“Going from single CPU to multi-core CPUs was a great challenge,” Huang says. “Going from multi-core CPUs to multi-node multi-core CPUs was an enormous challenge. And yet, when we when we create a CUDA, in our GPUs, we went from one processor core to a few, to now in the case of large scale systems, millions of processor cores, and how do you program such a computer across across multi-GPU multi-node, it’s a concept that that not easy to crack.
“And so I don’t really know how one programming approach or a simple API is going to make seven different type of weird things work together. And I can’t make it fit in my head. But programming isn’t as simple as a PowerPoint slide, I guess. And, you know, I think it’s just time will tell whether one programming approach could fit seven different types of processors when no time in history has it ever happened.”
Intel’s One API is a software programming model that will cross architectures and silicon, particularly useful for Intel, a multi-faceted company with fingers in many silicon pies. In an increasingly complex world of CPUs, FPGAs, GPUs, and accelerators, the need for overarching software able to tap into the many moving pieces in a modern data centre has increased exponentially. That’s what Intel’s attempting to produce with OneAPI and what Intel’s chief architect, Raja Koduri, has been keen to talk up to developers since it was unveiled officially at its Architecture Day last year.
Intel’s incorporation of a single, cross-industry programming language also flies in the face of single architecture proprietary languages, and that might rub Nvidia, and its CUDA platform, up the wrong way.
Nvidia’s proprietary CUDA platform is built to make the most of its GPUs, with API tools available to developers to squeeze all the performance out of its accelerators in GPGPU workloads. With datacentre and compute tasks favouring the parallel processing grunt of a GPU nowadays, especially for tasks such as AI, Nvidia’s CUDA API and libraries have become increasingly important to realising the company’s vision as a major datacentre player.
But there’s a positive for Intel in all of this shade throwing. Because if Jen-Hsun truly believes this has never been done before in history, it would be one hell of a victory for Chipzilla if it could pull it off. We’ll find out soon enough, OneAPI will head to develops in beta later this year.