CTO for Accelerated Computing
Steve Oberlin is the Chief Technology Officer for Accelerated Computing at NVIDIA.
Steve Oberlin’s large-scale computing technology career has spanned over 30 years, launched in 1980 at Cray Research bringing up CRAY-1 supercomputer systems. Starting in 1981, he worked for Seymour Cray as a designer and project engineer on the CRAY-2 and CRAY-3 supercomputers. In 1988, he led early massively parallel processing research at Cray that ultimately led to his role as the chief architect of the CRAY T3D MPP and its successor, the CRAY T3E. He holds 15 architecture and design patents for the T3D and T3E.
Steve was VP of Hardware Engineering in Chippewa Falls, Wisconsin, for Cray/SGI from 1996 until early 1999, responsible for hardware development and support of all Cray products and their follow-ons. Steve left SGI to found Unlimited Scale, Inc., in July of 2000, and spent the next 13 years creating new cloud computing infrastructure management and intelligent resource optimization technologies for start-up Cassatt and CA Technologies.
Steve returned to HPC at NVIDIA in November 2013. As CTO for Accelerated Computing, he is responsible NVIDIA’s Tesla roadmap and architecture.
HPCwire: Steve, congratulations on being named to our People to Watch list. In November, you commemorated five years as CTO of the Tesla business unit at Nvidia, responsible for the company’s flagship architecture and roadmap. It’s been a pinnacle period of success, where Nvidia has seen its datacenter business swell. What were some of the key milestones for the Tesla unit in 2018?
Steve Oberlin: Thank you, it’s an honor!
No kidding about this being a pinnacle period for us. It’s so great to see so much work by so many talented people paying off in mind-blowing achievements. There are too many milestones to list, but I’ll give you my top three:
First, there’s Oak Ridge’s Summit and Lawrence Livermore’s Sierra GPU-accelerated supercomputers topping the Top500 list. That, combined with so many other powerful GPU-accelerated systems around the world at their back, and dominating the top of the Green500 list, is a powerful validation of the scalability of our accelerated computing platform.
DGX-2, with NVSwitch and our Volta Tensor Cores, is the second. Nobody has built iron like this — such a crazy amount of performance and memory bandwidth in a shared memory server — since the days of giant vector supercomputers. Half a terabyte of HBM at over 15 TB/s, 125 TFLOPS double precision peak. It’s an awesome beast for HPC as well as AI, and a clue into how we’ll continue to scale application performance at Moore’s Law rate (or faster!), despite Moore’s Law’s decline.
The third has to be unprecedented application success, like both Gordon Bell Award winners achieving exascale and multi-exascale performance levels using Volta Tensor Cores on Summit, and seeing important workhorse applications like WRF being added to the 580+ accelerated applications already in our catalog, which now includes 15 of the top 15 HPC applications. This progress comes thanks in large part to PGI’s OpenACC being embraced by developers for its productivity, efficiency, and great performance. There’s a wave of OpenACC success stories emerging from hackathons around the world. It’s so fun to watch.
HPCwire: How would you describe the evolution of the Tesla GPU family under your leadership?
I’d say it’s hardly “leadership”, more like “cheerleading”. I’ve been privileged to work with some of the smartest and creative engineers in the world over my career and NVIDIA is extending that experience. They do all the rowing. I get to sit in the bow of the boat and pretend I’m steering.
There are two areas of optimization that are arguably the most important things we’ve done since CUDA made GPUs programmable. The first is NVLink/NVSwitch, which let’s us seamlessly scale the area of silicon a programmer sees as “one big GPU”. The second is the suite of optimizations in support of AI, including Tensor Cores and multi-precision arithmetic, which are now of course also being applied with such great results to HPC applications.
Together, these architecture features will continue to drive HPC application performance at the dizzy rate we’ve been growing for the past 5 years.
HPCwire: What is your perspective on the relationship and synergy between HPC and AI? What is the role of GPUs in this regard?
This is by far my favorite topic. There’s no doubt the current AI revolution — all of it, the self-driving cars, our smart phones actually acting smart, talking and listening to us, all the image and video recognition, classification, captioning, recommenders, AIs beating humans at games and serious tasks alike – happened largely because NVIDIA GPUs and CUDA were at a critical point of usefulness and wild price performance to enable deep learning algorithms to ingest and learn enormous new data sets. Almost every AI advancement in the past five years that required any significant compute horsepower was powered by NVIDIA GPUs. HPC + algorithms + big data = society-changing technology revolution.
Now, the AI revolution is coming back home to fundamentally alter how we do computational science.
Computational science is all about creating a mathematical model of the physics, chemistry, biology, or whatever phenomena you are investigating in software, and then programmatically executing it at sufficiently-interesting scale to make predictions that hopefully agree with and are relevant to the real world. But models are just models, and often – almost always – models necessarily take shortcuts to approximate various factors of computations that might be too computationally intensive or otherwise make the simulation too expensive and prevent simulating at meaningful scale or resolution.
Generically, deep neural networks are universal function approximators. They can learn a model from a well-designed and complete enough data set, which can be a high-fidelity, first-principles HPC algorithm. Once a model is trained, it can run in prediction mode, “inferencing”, at very high efficiency. It’s astounding. People are training AI models from simulations and real world data and producing models that are not only more accurate than traditional models, but can be several orders of magnitude faster.
This is going to redefine how supercomputing is done, and the architecture of supercomputers. AI supercomputing wants a universal architecture like ours, where the platform is simultaneously great at double-precision HPC, training AI, and inferencing at enormous scale. Summit and Sierra, and Japan’s new ABCI system are recent examples of this new kind of architecture.
HPCwire: What is Nvidia doing to further the democratization of AI?
What aren’t we doing? Even before my time, NVIDIA had a “CUDA everywhere” philosophy that propagated the same parallel architecture, programmable by the same languages and tools, on everything from 5-15 Watt mobile devices, to laptops, desktops and workstations, to the world’s largest and fastest supercomputer for the last 12 years. There are millions of GeForce game cards that can run all the frameworks. This whole AI revolution started with researchers training neural networks on their desktop GPUs.
Don’t have one or need more than you have? Volta with Tensor Cores, and a variety of other powerful NVIDIA GPUs, are in every cloud. I read the cost to train AlexNet is now down to just a few bucks. We’ve made a free repository of Docker containers with pre-installed configurations of every framework, optimized, accelerated, and maintained by NVIDIA’s performance experts that run on in-house or cloud resources.
If you’re an individual or small lab or business and don’t think you can get affordable access to NVIDIA GPUs to pursue AI, then you’re just not looking. It couldn’t be easier.
Philosophically, a platform and the community that stands on it is a democratizing force in itself. NVIDIA has invested a lot in the platform technologies and tools. Users benefit from those, but also from a network effect from being able to share and leverage work that is built upon that platform. A major benefit that’s definitely in the “democratization” vein is knowing that the great majority of models and codes in the thousands of GitHub AI repositories you may wish to leverage in your work were originally built on frameworks optimized for NVIDIA GPUs like your’s, and that as AI continues to explode, your platform is the vehicle for you to stay connected and leverage the advances being made by others at that leading edge, too.
In addition to the platform, we give a lot away to help people bootstrap to the next level, regardless where they are at. We’ve donated hundreds of the latest GPUs to researchers. We have Inception, a program that engages with AI startups (by the thousands, now), and DLI, our “Deep Learning Institute”, that offers a host of training and tutorial resources and classes. We have leading AI researchers in at NVIDIA and we collaborate around the world. We publish our results. We develop and maintain key technology like the CuDNN and NCCL libraries, which every DL framework leverages for optimal compute and communications. We’ve even released open source hardware, the design of our super-efficient inference accelerator, for anyone to include in their ASICs. Today, it’s “AI Everywhere” at NVIDIA.
HPCwire: Your tenure of innovation in HPC includes leading design work at Cray Research and SGI and now you’ve been driving Nvidia’s high-end architecture since 2013 – what excites you (most) about working in high-performance computing?
Well, I love hardware. I’ve always been a HW geek. When I joined NVIDIA, I was coming back from a decade away from HPC, and hadn’t really paid close attention to it. I had dinner with Bill Dally, NVIDIA’s Chief Scientist and a friend from the ‘90s when he helped us with the interconnect network architecture for Cray Research’s first MPPs, the T3D and T3E. When I worried my HPC knowledge might be stale, Bill told me, “Don’t worry, nothing’s really changed.”
That wasn’t quite true. GPU accelerated computing was new, and it was incredibly powerful. I had new perspective on the meaning of “future shock” the first time someone handed me a Pascal GPU. The Cray T3E was the first supercomputer in the world to sustain 1 TFLOPS on a real application. When I stepped away, there were still T3Es in production in the world, and, only a little over a decade later, I was holding its equivalent in my hand. That gave me goosebumps. (And, made me want to network a bunch of them together!)
The only thing that’s better than that, is the kick I get from seeing people take that amazing technology and do even more amazing things with it. I love nothing more than watching that never-before-revealed insight presented by excited researchers, the predictions confirmed or upended, the visualizations and interactions that are just impossible in the real world, and the fundamental impact on everyone on earth. HPC is important in ways that business or enterprise computing are not. Yes, our economy and the engines powering society are critical, but all of civilization is built on technology, and technology is built on science. It sounds corny, but it’s a thrill to know that the tools you’ve had a hand in creating are advancing civilization. There’s something deeply satisfying in that.
HPCwire: Outside of the professional sphere, what can you tell us about yourself – personal life, family, background, hobbies, etc.? Is there anything about you your colleagues might be surprised to learn?
Probably several things, but I’m only going to talk about one.
Back in the ‘80s, when computers had only alphanumeric displays and keyboards, my artist wife, Gwen, and I built a bunch of computer graphics hardware (frame buffers, floating point accelerators, and film recorders) of our own design and wrote various ray tracer, z-buffer, and scanline rendering programs, particle systems, paint programs, compositors, etc., trying to bootstrap a CGI business before there were such things outside Hollywood and Madison Ave. We did everything ourselves from scratch, wire-wrapping thousands of ICs, mining Siggraph proceedings and other papers for algorithms. It was completely fascinating. Even the failures were often visually spectacular. Of course, the desktops we were using were so feeble (even after I souped them up) it could take a day or more to ray trace a single 640×480 image, pixel by excruciating pixel, but it was unbelievably cool to get up in the morning and look at the newest image to emerge from this wild new viewport into a completely synthetic world.
It was nuts. We were terrible businesspeople who only liked to do the fun technical stuff, and it probably would have failed even if we’d done everything right, but it was about the most technical fun I’ve ever had (and I’ve had a lot of great technical fun). I was simultaneously working for Cray Research on the Cray-2 and Cray-3 during that glorious time, then effectively abandoned Gwen and doomed the graphics endeavor as I started to travel heavily in the early days of Cray’s MPP project. But, I still have a soft spot in my heart for computer graphics, the technology and the art, and still wonder what might have been if I’d taken the RGB pill instead of the blue one…
And the world moves on: 2018 brought even more future shock as NVIDIA demonstrates real time raytracing at ~10x higher resolution in professional and even gaming desktop GPUs. How cool is that?!