GTC 2015 Game Engine -like architecture to build advanced Scientific - PowerPoint PPT Presentation

GTC 2015 “Game Engine” -like architecture to build advanced Scientific Software applications mik@hue.no

In Oil & Gas since 2001 (also tested in Medical and CAE) Deployed by Oil Majors and Commercial Software vendors. Interpretation, Modeling & Processing & Imaging Reservoir Engineering Production Misc E&P Characterization This is a state-of-the- art product, not research…

3 Bringing the technical foundation of the gaming industry, to scientific and engineering software.

State-of-the-art visualization Visualization-driven interaction and compute, with a single processing pipeline for both. Big Data handling Accelerated compute framework Core Engine orchestrates visualization, compute and data. Developers access through cross- platform APIs.

HueStreams: 3D Wavelet compression / decompression 5 • Multi-CPU/GPU Parallelized 3D wavelet compression/decompression Technology Sequential read • Can be used for any volumetric data: seismic, medical, atmospheric, … . HDD 259 MB/sec • Numerically stable between CPUs and GPUs SSD 585 MB/sec • Auto-detects fastest decompression method at run-time (CPU or GPU) Flash PCI-e 2.7 GB/sec • Linear GPU scaling • On-demand parallel asynchronous data loading and production 1:15 compression GeForce 650m Intel i7 - 4 cores K6000/K40 4 x K40 example (old MacBook Pro) Compression 1.5 GB/sec TBA TBA TBA Decompression 2 GB/sec 3.5 GB/sec 10 GB/sec ~40 GB/sec Our 5X is lossless; but customers have been confortable with 15-30X compression.

6 GPU compression for interactive visualization and computation

Total dataset = 26 GB 1.1 Billion Tetrahedrons From “slow on a 16 nodes CPU cluster” to “fully interactive on K6000” On the fly interactive streamlines computation and new properties calculations, etc…

8 CFD FD example: mple: • 2 Formula1 cars with full CFD simulation (pressure, velocity and velocity magnitude properties). • Total dataset = 26 GB = 1.1 Billion Tetrahedrons • Simulation model: – 5.1 GB Vertices – 7.4 GB Properties data – 16 GB Topology/Geometry • Traditional: 32-64 GB (2-4X blow up for rendering/algorithm structures) • HueSpace: 2.3 GB (7X structural optimization, used directly for rendering and algorithms) Over 14-28 X DATA SIZE OPTIMIZATION Scale out for geometry data with multi-GPU single system: 12GB GPU ≈ over 1 Billion cells 16 GPUs ≈ over 16 Billion cells

Seismic Demo 9 Amazon AWS G2 instance w /GRID K520 1536 cores + 4GB GDDR5 Calgary Scientific PureWeb-enabled demo • Public data from Australia • 24,000 Km 2 ( almost the size of Switzerland or Massachusetts ) • 7 seismic surveys • 1.5 TB combined

11 “We cannot simply if-then-else out of this…” Jen-Hsun Huang NVIDIA’s CEO