1 Supercomputing: Yesterday, Today &Tomorrow Peter Ungaro | CEO, Cray 2017 RICE UNIVERSITY O&G HPC CONFERENCE Very excited to be here today– had the pleasure of presenting here over the last 10 years (about 5 years ago) I was asked to take a look back at the last 10 years, take a snapshot of today and then look ahead 10 more years. ….you know, as I’ve gotten older I tend to think that 10 years really isn’t that long! My kids think of a decade as an eternity!

2 will not get us where we needWhat got us here today will not get us where we need to be tomorrow A transition is occurring in the industry to enable more efficient discovery and recovery – driven by economics, technology, and innovation. At the same time, supercomputing is being redefined. So while many things remain the same…many things are changing in the world of supercomputing as it relates to earth systems modeling in O&G Industry leaders are recognizing that approaches that got them here today, are not going to get them where they need to go.

3 Oil & Gas: Yesterday One shot per node computing model New survey techniques and technologies are starting to drive model size and complexity The adoption of approaches such as RTM, FWI, etc. have been driving the need for more capability Processing system deployments have gone from teraflops to petaflops As we look back, the O&G industry has gone through significant change over the past 10 years. New survey techniques and technologies are driving model size and complexity: PGS’ Triton Gulf of Mexico survey collected in 2013/14 Adoption of approaches such as RTM, FWI, etc. have been driving the need for more capability: processors, interconnect, storage Data sets grew but not fast enough to force changes from the “one shot per node” computing model Processing system deployments have gone from teraflops to petaflops: Mainstream adoption of RTM variants and production usage of FWI have driven seismic processing systems over the petaflop boundary

4 Supercomputing: YesterdayProcessor clock rates drove performance gains Focus on driving the cost out of computing And when we look back at computing….stayed in the same operating model of one shot per node model and successfully drove out the cost The use of dominant network technologies like Ethernet and InfiniBand and commodity software stacks handcuff the scalability and efficiency of parallel HPC architectures Few systems were engineered for extreme scalability

5 Oil & Gas: Today Data sets growing from terabytes to petabytes The impact of new survey techniques and technologies are driving new geoscience and new workflows Considering the transition from a “lowest cost” to a “competitive edge+ROI” perspective System architects actively are planning for double and triple-digit petaflop systems The O&G industry is at a transition point today. New survey techniques and technologies: The results created from new surveys are driving new R&D to provide clearer images and The size and complexity of newer surveys is driving R&D to provide more efficient workflows Data sets growing from TB to PB: New survey techniques have increased the data sets by 1-2 orders of magnitude and new algorithmic approaches forcing the storage of intermediate results during a workflow Bigger systems, utilizing significant parallelism: Work is being done now to size and scope next-generation supercomputing environments -- Total work at ORNL(Cray Titan) to test RTM at scale: Processed the 2,793 Shot profiles of SEAM Model over 18,508 GPUS in 54 minutes

6 System throughput is the major bottleneckSupercomputing: Today Parallelism is broadly recognized as mandatory to achieve processing efficiency Processor clock rates have stalled System throughput is the major bottleneck Major technology shifts have impacted systems and software architectures – system capability & throughput is King (not feeds and speeds) Serial processor speed ups have slowed or halted – clock rates have stalled -- Manycore/Multicore architectures have gone mainstream and showing value in some O&G codes (KNL, GPU’s) Memory bandwidth has become a major bottleneck – on package memory Power has become a major limiting factor New interconnect topologies have gone mainstream: Dragonfly – advantages at only 32 nodes – congestion mgmt: adaptive routing, job placement insensitivity, HW collectives Parallelism has become mandatory if processing efficiency is to be achieved – optimized libraries (MPI +30%; Sci Libs +40%; MPI I/O +7X)

7 OIL & GAS TOMORROW Will require an order of magnitude more processing power in the next few years (approaching 100 PF) Reservoir simulation will also become capable of handling realistic earth models Another order of magnitude more processing power on top of that within 10 years (approaching 1 EF) The convergence of HPC and analytics will present new possibilities to optimize O&G workflows and quality of results As we move the clock forward 10 years, our biotech friends will figure out how to let me regrow my hair again, but more importantly… The efficient utilization of realistic earth models will require machines with two orders of magnitude more processing power than the typical systems in production today. This statement is based on sizing estimates from a number of our customers. Datasets in the multi-PB’s.

8 TOMORROW SUPERCOMPUTING Parallelism is the new normalMemory and storage hierarchies will continue to get deeper Applications architectures will change to harness the power of “wider” computers Processor choice will grow with heterogeneous systems Interconnects must improve rapidly to deal with congestion & throughput Current processor technology (CMOS) will eventually run out of steam Parallelism is mandatory – software engineering is king –Significant R&D commitment will be necessary to eliminate bottlenecks that hinder the scalability of new parallel codes – will further increase the importance of tools and libraries, and expertise in parallelization Applications architectures will change to harness the power of “wider” computers – Cray expertise, programming environment, and libraries enable this change New interconnects with minimal hop counts and scalability to hundreds of thousands of nodes will appear and provide improved system efficiency – mandatory for exascale systems Memory/storage hierarchy will become deeper – thus increasing application complexity, but dramatically improving performance Processor choice will grow – Systems will likely be heterogeneous, with additional processors to choose from – Intel, NVIDIA, AMD, ARM, FPGA, others Current processor technology (CMOS) will run out of steam – what’s next?

9 Exascale Will Be the End of the CMOS Era…Low diameter networks with optics O(100k) nodes, ~30 MW O(10-100M) system threads NVM for storage cache Memory system on package Disk for archive Flash for main storage 10-50 TF per node Wide vectors (or GPUs) So let’s take a look at what this might look like in a system a few years from now… CMOS has 10+ years left of life before it hits fundamental limits. Exascale is the last factor of 1000 we’ll see with this technology. Can connect up these massive systems with a very efficient, low-diameter network that takes only a single optical hop! First exascale nodes might be 10-20TF; could perhaps increase to 50TF over time M total threads in the system. Memory will move on package. ~10x better bandwidth and power efficiency than external memory – some systems with no DRAM Storage hierarchy will take advantage of new media…

10 Analytics Models Scientific ModelsData-Intensive Processing Analytics Models Scientific Models Our goal is to help you to “Model the World” Data explosion driving unprecedented abilities – fusion between supercomputing and big & fast data Help our customers compute/store/analyze in this environment Even do operational analytics on the supercomputer itself: increase reliability, diagnosability, performance, etc… Compute Store Analyze

11 In-memory processing of 600+ TB of dataA Clearer Image FWI + In-memory processing of 600+ TB of data A great example of this is from our friends at PGS…. A huge survey (Triton) in the Gulf of Mexico with 2.6 million km (common midpoints) acquired Full Waveform Inversion for the highest resolution, highest fidelity subsurface image The first question is if there is a better way to process the data – one shot on multiple nodes using globally addressable shared memory and parallelism

12 “Ground truth” BenchmarkA SMARTER MACHINE “Ground truth” Benchmark Conventional FWI With Machine Learning Abel: With the help of research from Prof. Marten de Hoop at Rice University, PGS applied ML optimization techniques such as regularization and steering to determine the velocity model of a well-established, challenging benchmark synthetic data set — the BP velocity estimation benchmark.  Standard benchmark/actual subsurface FWI with manual velocity model determination and tuning FWI with ML iterations to converge on the best velocity model --finding similar results in local weather forecasting…

13 RTM problem size approaches 200 GB per shotNew approach for Reverse Time Migration (RTM) applied to 660 TB GOM survey achieved 129 M traces/minute First “Trillion Cell” Reservoir Model at the King Abdullah University of Science and Technology(KAUST) Breakthrough in Parallel Reservoir Simulations ,800 processors utilized at NCSA SUPERCOMPUTING VISIONARIES… TODAY Not everyone is waiting on tomorrow… - ExxonMobil -- Breakthrough in Parallel Reservoir Simulations - Blue Waters - 716,800 processors utilized at NCSA to make better investment decisions by efficiently predicting reservoir performance in less time – “1,000’s of times faster than a typical res sim model” - Saudi Aramco -- Largest reservoir simulation in history – Shaheen II – 1 trillion cell model versus typical models used in industry today of 1-20 million cells (nearly 1PB of data) - PGS -- Most complex seismic survey ever recorded at the time – Processed on Abel - 660TB survey: RTM approaching 200GB/shot and achieving 129 M traces/minute So while what got us here today will not get us where we need to be tomorrow…I’m convinced that we have a pathway forward that is a different, but for the leaders of the industry is one that can successfully be taken over the next decade!

14 THANK YOU Pete Ungaro | CEO, Cray2017 RICE UNIVERSITY O&G HPC CONFERENCE So while what got us here today will not get us where we need to be tomorrow…I’m convinced that we have a pathway forward that is a different, but for the leaders of the industry is one that can successfully be taken over the next decade! Thank You