SMT thrashes Moore’s Law (July 2005)

SMT thrashes Moore’s Law (July 2005)

SMT’s SURE reservoir flow simulation unit has achieved blistering performance using graphics cards for number crunching. The technology is set to outpace Moore’s Law by orders of magnitude.

There will be no need to wait on Moore’s Law to align CPU speeds with your reservoir simulation requirements if technology under development by Seismic Micro Technology’s Sure reservoir simulation unit comes good. The idea is simple, instead of using the computer’s CPU for number crunching, use the graphics processing unit (GPU) of a gaming card.

Smith

SMT president Tom Smith told Oil IT Journal, ‘You’ll wet your pants when you see the spectacular performance the GPU brings. A top of the range computer CPU produces around 4 GigaFlops (billion floating point operations). But today’s GPUs can already run at around 30 GFlops. Performance is fast approaching the Teraflop—truly a supercomputer on a chip!’

Ganzer

SMT’s Leo Ganzer presented the technique at the Madrid EAGE last month, showing how reservoir simulation can now be performed in an interactive mode—as opposed to traditional batch processing. Interactivity is facilitated by making the simulator a part of the visualization application.

Moore beat

GPUs have evolved from fixed-function devices into programmable chips with high floating point operation counts and a competitive price/performance ratio. Current GPUs outperform CPUs on certain computations. But whereas ‘Moore’s Law’ is now slowing to a CPU speed doubling every 18 months, GPU speed is currently doubling every six months!

NVidia

The simulator was developed on dual NVidia GeForce cards using NVidia’s C for graphics (Cg) language. Grid block properties and transmissibilities are pre-computed and stored in GPU textures. The program reads all input from the grid using texturing operations, calculates the coefficients, solves the equations and updates the dynamic properties for the new time step. At the end of a render pass, textures are updated via a copy from the frame buffer to the texture. Between steps, interaction with the model is possible by changing well production/injection rates or the time step sizes during the simulation—using slider bars for input.

Gpgpu.org

GPUs are highly parallel in nature and because the time between new generations of GPUs is currently much less than for CPUs, it is anticipated that the technology has enormous future potential. The technique leverages open source code for GPU processing from the General Purpose Computation Using Graphics Hardware Forum. More from www.gpgpu.org.

Click here to comment on this article
If your browser does not work with the MailTo button, send mail to info@oilit.com with Comment_OITJ_0507_1 as the subject. Copyright © 2005 The Data Room - all rights reserved.