So now you know what GPGPU is, what it is not, the magnitude of the performance gains you can expect to get now and in the near future, and a bit of history as to how we got here. But how can you effect the necessary change in your organization to benefit from this new technology?
Well take it from me… It ain’t easy. I’ve been following GPGPU since Nvidia released their first CUDA SDK. Given that I currently manage several computational grids for my firm, I was quick to see the potential benefits. I naively assumed that all I would need to do to get my firm moving in the GPGPU direction was to build some of the Nvidia CUDA option valuation examples and integrate them into our grid. Once I showed people the potential gains that could be achieved they would happily start moving down the path to GPGPU nirvana.
Boy was I “mistaken”. The first complaint was, “It doesn’t support double precision so it’s useless to us. Get back to me when it supports doubles.”. Six months later, after it supported double precision the complaint was, “This requires me to rewrite my code so I’m not interested. Get back to me when I can just recompile my code and it is magically 100X faster.”. Well that silver bullet may come sometime down the road but I’m not going to hold my breath (and you shouldn't either).
The most frequent reaction was, “I already have the computational capacity to get my work done. Why should I spend the necessary development resources just to save a few million dollars?”. Which to people outside of the financial industry this must seem like an odd reaction. Why wouldn’t you spend 6 months of development time to save a few million? Well when you take into consideration that many of these systems make millions of dollars a day, any savings could be wiped out (and then some) by a single mistake in rolling out GPGPU based versions of the models.
With all of these impediments to change how do you get the ball rolling? Well you have got to get involved in the initial stages of something new. Get involved when they don’t already have a system. Get involved before they spend $10,000,000 for hardware. The resistance to change seems to disappear when you can show them their model running on a GPU side by side with a CPU based version. The resistance will disappear when you show them that they can get a 30X – 100X speed up over their CPU based model. Point out that instead of paying $10,000,000 for hardware they can do the same number of calculations for $250,000.
This concludes my series on GPGPU Overview. I hope you found it insightful. The example sections will walk you through some CUDA, OpenCL, and PyCUDA code to get your feet wet. Under products you will find useful information on current offerings from many different vendors (hardware and software). The Resource section will point you to documentation and tutorials to make you productive quickly. News will keep you up to date on various GPGPU topics.
If you would like to follow my site, be alerted to updated information, and hear about interesting events in the GPGPU industry just go to my News section and subscribe to my site feed.
Live Free or Die