Speaking at the 2015 Nvidia GPU technology conference earlier this year, Dominic Walsh explained how Schlumberger is introducing GPUs into its ‘next generation’ Intersect reservoir simulator. Since 2011, clients’ model sizes have risen from a few tens of million cells to as many as 2 billion cell models running on Linux clusters. Adding surface facilities, economics and uncertainty makes for ‘compute bound’ models, hence the interest in the potential of the graphics processing unit (GPU) to accelerate simulations. Over the next 5-10 years Schlumberger plans to migrate its user base from Eclipse to the Intersect ‘high fidelity’ simulator and is keen to add new GPU hardware to the mix. Parallelizing computations to the GPU is complex. Modeling in the vicinity of the well bore is ‘too small and complicated’ and will remain on the GPU.
Walsh’s talk was co-authored by Stone Ridge Technology whose Dave Dembeck presented on how GPU acceleration has been leveraged in its own Echelon reservoir simulator. Echelon was developed from the ground up for GPUs and claims ‘10x to 50x faster’ runtimes than leading commercial simulators. Dembeck observed that as simulations speed up, they can create a ‘downstream bottleneck’ as engineers are inundated by model results. Workflows group results into ensembles and visual tools help scenario and model triage.
Saber Feki presented KAUST’s experience with OpenACC, a higher level, cross platform route to parallelizing scientific applications to GPUs and other hardware. KAUST showed OpenACC ports of seismic migration and inversion codes for use on its Shaheen II Cray XC40, a 100 tonne monster with a 7.2 petaflop peak capacity. Shaheen II uses 6,192 x 16 core Intel Haswell CPUs. A GPU upgrade is planned as the machine has maxed-out on power usage at 2.8 megawatts!
Our ‘best pics of the show’ award goes to Michael Heck (FEI - Visualization Sciences Group) who showed how Schlumberger’s Petrel leverages Nvidia GPUs via FEI’s Open Inventor toolkit. Volume rendering and interaction with 220 giga voxel datasets is possible. Heck reported that the latest Quadro K6000 is around twice as fast as the K5000 with same software. Resist
Finally Jonathan Marbach’s (CGG Software) presentation on GPU acceleration of acquisition footprint removal in post-stack seismic data is replete with benchmarks and offers a good history of GPU evolution.
© Oil IT Journal - all rights reserved.