You might not know it by looking at the consumer applications out there, but general-purpose GPU computing does have uses other than video transcoding. No, really. Wall Street & Technology has written a story about one good example: Bloomberg's latest server farm, which uses Nvidia Tesla cards to run pricing operations at a fraction of the cost of traditional servers.
According to the story, Bloomberg has been working out prices for "1.3 million hard-to-price asset-backed securities such as collateralized mortgage obligations (including cash flows, key rate duration and such)" every night since 1996. Four years ago, the company developed a more precise yet more computationally demanding model, running it only for clients who asked.
Demand grew. Eventually, Bloomberg worked out that calculating all results through that model overnight would require a tenfold increase in processor cores, from 800 to 8,000. Rather than take the brute-force approach, the company went with Tesla GPUs. Now, instead of 1,000 servers with eight cores each, Bloomberg is able to run its calculations on just "48 server/GPU pairs."
The code did take a year to port—presumably using Nvidia's C for CUDA programming interface—and some parts still have to run on x86 processors. However, Wall Street & Technology says the GPUs still do about 90% of the work and deliver a performance increase of about 800%.