i do! i write custom code for a small "consultancy" in geophysics (we're a handful of people from academia that somehow ended up developing software for people - and it doesn't really have to be geophysics, that's just where we started...).
my most successful project was an inverse problem for mining - taking a bunch of seismic traces and, from that, calculating a 3d map of the source structure (microseismicity - this technique is called SET (seismic emission tomography)). there's an abstract here - http://adsabs.harvard.edu/abs/2010AGUFMIN41A1352D (if anyone wants i can send them the poster, if i can find it!).
incidentally, the AMD OpenCL implementation worked fine on Intel processors already (and has done for a year or so now). i made some notes at http://www.acooke.org/cute/Developing0.html
it's very similar to cuda, but has a couple of advantages. first, it's really well supported on macs. second you aren't stuck with nvidia's hardware (their support is terrible - had a client stuck waiting on results for weeks because they wouldn't respond to issues that eventually turned out to be a failed board). the main disadvantage is that the associated tools/libs aren't as mature.
and a final piece of advice - if you have a finite budget, and don't need 64bit support, you get way better performance per $ using average gaming cards rather than their top-end GPU hardware (although i suspect we'll try out amazon's gpu cloud stuff soon and avoid our own hardware for projects that don't have huge data sets).
AFAIU, OpenCL is still a new technology - poised as competition to Nvidia's rather popular CUDA. However, in contrast with CUDA which is Nvidia-specific, OpenCL is an open standard backed by many vendors (Intel, AMD, Nvidia, Apple, IBM, Qualcomm and others).
Don't for get that GPGPU is becoming hotter and hotter now that the computing devices blend and become heterogeneous (i.e. CPUs come with embedded GPUs, GPUs doing general-purpose computations, special acceleration hardware, etc). OpenCL is supposed to help with programming these devices.
One of the things they are trying to GPU accelerate is improved MRI scannings - imagine taking images for every 1mm and then stitch it together so that you get a 3d model of the area which the doctors can look at from every angle.
Unfortunately that takes way, way too much computer power to be feasible today.
Guessing here but so they can be turned into a polygon model for easier display, yes you can directly volume render the data but say a 1024 x 1024 x 1024 cube even at a byte per voxel would be a large amount of memory, most of which will be empty
I'm guessing OpenCL is being used to accelerate Marching cubes or some other similar algorithm
Some folks were considering using CUDA from the Linux kernel. I imagined it being used for accelerating block hash calculation in block-deduplicating filesystems (that would be a nice addition to BtrFS, BTW).
And, if you can use CUDA to accelerate stuff like that, you can use OpenCL.
my most successful project was an inverse problem for mining - taking a bunch of seismic traces and, from that, calculating a 3d map of the source structure (microseismicity - this technique is called SET (seismic emission tomography)). there's an abstract here - http://adsabs.harvard.edu/abs/2010AGUFMIN41A1352D (if anyone wants i can send them the poster, if i can find it!).
incidentally, the AMD OpenCL implementation worked fine on Intel processors already (and has done for a year or so now). i made some notes at http://www.acooke.org/cute/Developing0.html
it's very similar to cuda, but has a couple of advantages. first, it's really well supported on macs. second you aren't stuck with nvidia's hardware (their support is terrible - had a client stuck waiting on results for weeks because they wouldn't respond to issues that eventually turned out to be a failed board). the main disadvantage is that the associated tools/libs aren't as mature.
and a final piece of advice - if you have a finite budget, and don't need 64bit support, you get way better performance per $ using average gaming cards rather than their top-end GPU hardware (although i suspect we'll try out amazon's gpu cloud stuff soon and avoid our own hardware for projects that don't have huge data sets).