In her position as emerging market and technology director at Cray, Arti Garg doesn’t just have a front-row seat to the future of computing, she plays an active role in making that future happen. Key to Garg’s role is understanding how deep learning scientists are using state-of-the-art HPC infrastructures and figuring out how to push those limits further. A couple months ago at the GPU Technology Conference, HPCwire sat down for a conversation with Arti Garg. Here we share a lightly edited version of that interview.
HPCwire: Can you describe your role at Cray and where you sit in the company?
Arti Garg: I’m the Emerging Market and Technology Director at Cray and am focused on how supercomputing capabilities are relevant to AI. My role is to see around the corner – what’s next for AI as it relates to high-performance computing – and assess where our systems add value in the AI market. I’m focused on fostering the right partnerships and meeting the right product requirements to ensure our customers are equipped with the best tools for their big data workloads.
HPCwire: You were brought into the AI team about a year ago – what are some of the key achievements during the time?
Garg: One of the first initiatives I worked on was a substantial overhaul to our Accel AI reference configurations that we announced at ISC last year. Accel AI reference configurations help make it simpler for customers to understand how to design their own AI systems using Cray products. I worked with members of the AI team to determine what the configurations should look like and what problems they solve for our customers. We developed the updated configurations in response to our customers expressing the need for systems that addressed their entire AI workflow. They need systems that have the right nodes for data preparation and the right nodes for model training. So, we put in a heterogeneous mix of compute nodes and hybrid storage–flash and disk storage–to make it more cost-effective and we packaged it with our Urika-CS AI and Analytics software platform.
HPCwire: How does your department and the goals of your department fit with the overall goals of the company? How do you help execute on those?
Garg: AI is of strategic importance to Cray so we are very closely aligned with the overall goals of the company. I spend a lot of time exploring what’s new in the space, but at the end of the day we want to be able to deliver the optimal systems for our customers’ needs. We bring a valuable perspective as a systems company, and our goal is to deliver systems that help customers can run their AI workloads more effectively and more efficiently.
Compared to traditional HPC workloads, what’s different for Cray in the AI world is the expanded meaning of performance. It’s not just about the system being big and fast–there’s additional dimensionality to how you define as performance. For example, we know that data scientists often want to be able to fluidly move data back-and-forth from a data preparation workload stage into a model training stage. That’s not exactly just a scale problem or just a speed problem. But these are the types of challenges that might trigger organizations to approach Cray. It can be a systems problem and a speed problem and a bandwidth problem, etc. It is often a problem that leverages many of our core capabilities.
HPCwire: Do you have an example use case that would further illustrate this point?
Garg: Developing models for autonomous vehicles is a great example. These models are built using several different data types, including video and LIDAR from vehicles. Initially, the focus from an HPC perspective was on dense GPU compute nodes for training the deep learning models that form the basis for the vehicle intelligence. Increasingly, though, we are seeing customers focusing more attention on storage needs for the massive amounts of data they are collecting and using to develop these models. I anticipate that storage will become even more important as the regulatory landscape around these models matures, since compliance will likely include archiving the specific processed versions of data used to train deployed models. The latter points to another shift, which is that we are increasingly hearing that customers want the ability to easily test out several different ways of processing their data to determine the impact on model outcomes. Since data processing is often run on CPUs, that drives the need for a mix of node types on the same system–all with the ability to talk to the same storage.
HPCwire: We’ve been using the term AI, which has some definitional challenges, so it would be good to get clarification. What are you including under AI rubric?
Garg: That’s a great question and an important one to clarify. We think of AI as the broad category of “computer systems that mimic or augment human cognition.” This includes everything from analytics to deep learning. Practically, we see the most interest in HPC systems for deep learning due to the compute-intensity of model training. But this is starting to change as customers’ AI workloads move from early-stage prototypes to larger production workloads. Increasingly, we are finding that AI systems need to address all areas of AI, including analytics and statistical machine learning that might be part of a data preparation stage.
HPCwire: Increasingly, HPC systems are being built to support machine learning and analytics workloads as well as traditional mod-sim. Does your group support these efforts? Take for example the Berkeley Lab computing facility, NERSC, which is home to Cori – a Cray XC40 supercomputer, and future home of Perlmutter, the first announced Cray Shasta system, set to debut in 2020.
Garg: Yes, absolutely. Machine learning is being explored or adopted across our customer base. In many cases they have acquired systems specifically for AI workloads, but increasingly we see traditional mod-sim being augmented with machine learning and analytics.
With NERSC, specifically, we have a collaboration to work jointly to explore AI use cases at scale and the necessary ecosystem to support these use cases. The CosmoFlow project that we did last year is a good example of our collaboration. We believe CosmoFlow is the first large-scale science application to use the TensorFlow framework for synchronous training at supercomputing scale. However, CosmoFlow didn’t just run out-of-the-box. We had to make several adjustments to be able to run data parallel training across 8000+ nodes. Separate from CosmoFlow, we are working with them on some interesting AI workloads that may soon find their way onto Cori, and I think it’s pretty cool for the DOE’s science mission.
HPCwire: Coming more from the data science side, what’s your perspective on HPC, where it’s been and where it’s heading?
Garg: It’s interesting – my background is in astrophysics – so I’m from a slightly earlier generation of scientists working with large data sets. We had very large computing workloads, but they were dominated by data processing. We processed our data in a way that we could run on large commodity clusters, despite many of the members of my collaboration having access to supercomputing systems. One of the reasons we did this was that the user model at those centers was typically not designed for long-term storage of large data sets. One of the needs that Cray Shasta was designed for – and what you are going to see in some of these upcoming national lab procurements – is the convergence of the need to run traditional HPC workloads like simulation as well as new workloads like AI to drive scientific progress. An emerging challenge is the extent to which our customers need to really grapple with the size of data that’s being produced from scientific experiments, which require specialized computing.
From the HPC + AI perspective, we’re at the frontiers of science right now, not just computing science in particular. I believe that is what’s really important. Going back to the DOE science mission, there are multi-billion-dollar investments into scientific facilities from the DOE and from many other research agencies across the world. They are producing these streams of data that today’s HPC systems couldn’t keep up with if they were to process every datum that came off of them. It’s a very exciting time!
HPCwire: So you have the architectural change and then there’s the bringing the AI algorithms into workflow to offload that and then you can do more traditional mod/sim with the HPC …
Garg: If you think about simulation workloads, if there are parts of a simulation that aren’t so important to the outcome, you may consider using deep learning to interpolate across that part rather than use compute cycles to get the details.
HPCwire: Do you have any thoughts about the heterogeneous datacenter? I was just at a press briefing with Jensen (Huang, Nvidia CEO) who noted the Top500 has lots of heterogeneous or accelerated systems – but in the broader market that’s not the case (as of yet).
Garg: That’s what the Shasta platform was designed to anticipate: this notion that people will run many different workloads and they are going to need a variety of types of computing infrastructure to do so. That’s what make our new Slingshot interconnect so exciting. It recognizes that not everything is about speed and scale, sometimes it is about being able to connect to the right storage. One of the design parameters of Slingshot is Ethernet compatibility so Shasta systems can be fully interoperable with standard datacenters and connect to existing network storage systems. This design aspect is going to be critical to helping users –commercial users, government, scientists, etc. –be able to run all of their workloads.
HPCwire: What are you looking forward to? What are the goal posts you are looking forward to this year or in the next few years?
Garg: This goes back to your question about how AI is evolving. It’s evolving away from people just testing it out, like a new toy, to being integrated into what people are doing, whether that be scientists who are changing the way they approach their scientific research or whether that be companies that are changing processes because they have new technologies available to them. Cryogenic electron microscopy, or cryo-EM, is an example of an application that is transforming drug discovery in the pharmaceutical industry. We are interested in this technology, because processing the large volume of data produced by the microscope is a major driver of compute demand.
My job is to see around the corner, but the truth is this is a rapidly changing landscape. The through-line from what we see today to what will be tomorrow is not straight. Pure computer science research is changing. People are developing new ways to build models and new model architectures. There are all these new technologies: new computer hardware and compute infrastructure and new ways of programming and architecting software. And there are other technologies, like IoT sensors, driving massive data generation. So, the fundamental question is how we will take all of this and put it together to do something useful. If history is a guide, it’s not always logical or predictable what technology will win or gain traction first. I’m interested to see how all of this shakes out.