The Convergence of HPC and Deep Learning
Download as pptx, pdf2 likes2,459 views
The document discusses the convergence of high-performance computing (HPC) and deep learning. It notes that GPUs, originally developed for HPC, now power advances in deep learning for applications like image recognition. Deep learning is also being applied to HPC domains to complement simulation methods. The speaker from NVIDIA outlines their work developing systems like the Legion programming framework that can handle both HPC and deep learning workloads, as well as research toward building exascale machines capable of both.
More Related Content
What's hot (20)
Similar to The Convergence of HPC and Deep Learning (20)
More from NVIDIA (20)
Recently uploaded (20)
The Convergence of HPC and Deep Learning
- 1. The Convergence of HPC and Deep Learning Bill Dally, SC16
- 2. AN OVERVIEW… 1. The Revolution in AI 2. Synergy of Deep Learning and HPC 3. Capabilities of Handling Deep Learning and HPC 4. NVIDIA’s work in HPC and Deep Learning 5. Concluding Thoughts
- 3. • 2006: Launched CUDA at Supercomputing • 2008: NVIDIA first Top 500 system • 2009: Designed Fermi as a high performance computing GPU • 2013: Andrew Ng and Bryan Catanzaro work together on deep brain project running GPUs • 2016: Created the NVIDIA SATURNV, showcasing NVIDIA’s capability as a system vendor and took #1 spot on Green 500 list Image Source: NVIDIA The Revolution in AI Content Source: Bill Dally, SC16 Talk
- 4. TODAY, PEOPLE WHO DISCOVER THE BEST SCIENCE ARE THE PEOPLE WITH THE BIGGEST SUPERCOMPUTERS
- 5. The Revolution in AI | Supercomputing Science is being enabled by supercomputing, whether it’s climate science, combustion science, or understanding the fundamentals of how the human body works to develop more medications. Image Source: NVIDIA What’s exciting is that the same technology enabling this powerful science is also enabling the revolution in deep learning, and it’s all enabled by GPUs. Content Source: Bill Dally, SC16 Talk
- 6. The Revolution in AI | Big Data Last year, a deep neural network defeated one of the best human players in a game of ‘Go.’ This is a game with an enormous optimization space. There’s no way to search over all possible combinations. The graph below, shown by Jeff Dean a year earlier, highlights the number of individual projects at Google that use Deep Learning. Content Source: Bill Dally, SC16 Talk
- 7. There is an interesting synergy between deep learning and HPC. The technology originally developed for HPC has enabled deep learning, and deep learning is enabling many usages in science. For example, it’s good at recognizing images and providing classification. Synergy of Deep Learning and HPC Content Source: Bill Dally, SC16 Talk
- 8. Synergy of Deep Learning and HPC Deep Learning can also apply to more traditional HPC applications. These applications can use the deep network to learn by taking a lot of the cases that have been simulated and training the network to identify similar cases. Then take a new case, feed it into the deep network, and it will predict what the output will be. Content Source: Bill Dally, SC16 Talk
- 9. Synergy of Deep Learning and HPC Both need arithmetic performance and performance tends to be judged in terms of performance per watt. All of our machines are constrained by a fixed number of watts whether it’s deep learning or HPC. Content Source: Bill Dally, SC16 Talk
- 10. Differences between HPC and Deep Learning There are some differences, but they’re small. If the machines are built and provisioned in the right way, then one machine can meet both. For HPC double precision 64-bits of floating point arithmetic is needed to get numerically stable solutions to a lot of problems. For deep learning training, you can get by with 32 bits. In addition, Deep Learning needs more memory per flops. But it’s just a question of how to provision that memory. HPC is more demanding of the network bandwidth, deep learning less so. Content Source: Bill Dally, SC16 Talk
- 11. Capabilities of Handling HPC and Deep Learning The HPC market is not big enough to fund the billion dollar a year investment it takes to develop chips like Pascal. So it’s not sustainable to build a chip just for HPC What’s great about GPUs is that they have many successful markets that have convergent requirements. Content Source: Bill Dally, SC16 Talk
- 12. Our Work in HPC and Deep Learning We are working in collaboration with a number of the national laboratories and with Stanford University on a system called the Legion programming system, which is an example of what I call target independent programming. With target independent programming, the programmer does what they’re good at, which is describing all of the parallelism in the program, not just how much to exploit. Image Source Content Source: Bill Dally, SC16 Talk
- 13. Our Work in HPC and Deep Learning Using this data model, which is what distinguishes this from a lot of the other task-based runtimes, it maps it onto a system in a way that maximizes use of the memory hierarchy and the use of the compute resources that can remap from one machine to another quickly. Image Source Content Source: Bill Dally, SC16 Talk
- 15. Concluding Thoughts It’s really exciting watching this deep learning revolution going on because it is very synergistic with HPC. They need the same things and building solutions for HPC map exactly right for deep learning. The deep learning techniques get turned around and are applied to predictive methods that complement the simulation methods being used for scientific things also for automatically analyzing data sets. Content Source: Bill Dally, SC16 Talk
- 16. There are some gaps left, but I’m confident that if we continue plugging away at some of the research lines we’re looking at, that we will be able to build an exascale machine at something close to 20 megawatts in 2023, if not sooner. GPUs are viable not just for HPC but also for deep learning and graphics. We have an economic model that works. We can sustain the engineering effort needed to bring you a new GPU every generation. Concluding Thoughts Content Source: Bill Dally, SC16 Talk
- 17. About the Speaker: Bill Dally Bill Dally joined NVIDIA in January 2009 as chief scientist, after spending 12 years at Stanford University, where he was chairman of the computer science department. He has published over 200 papers, holds over 50 issued patents, and is an author of two textbooks. Dally received a bachelor's degree in Electrical Engineering from Virginia Tech, a master’s in Electrical Engineering from Stanford University and a Ph.D. in Computer Science from CalTech. He is a cofounder of Velio Communications and Stream Processors. FOR THE FULL RECORDING: WATCH HERE
- 18. LEARN MORE ABOUT THE INTERSECTION OF AI AND HPC INSIDEBIGDATA GUIDE