Sekerob,

Are you factoring in HFCC restarting? Or a new project? Next college year is around the beginning of Sept, right? I project I need another 6 months at my current rate to reach my goal on HCC. If it really is going to finish that much faster, I will make the necessary adjustments so that I can reach 50 yrs on HCC (only 20 now).

I've seen some reasons for stopping efforts for a GPU version at various BOINC projects, but no reason why it can never be done.

At Rosetta@Home, they decided that the algorithm of their current application is so serial in nature that they had two choices:

1. Put as many copies of that algorithm into the GPU as it can find enough graphics memory for. Since each copy of that algorithm currently requires about 600 MB of memory to run, a fairly small fraction of the GPU boards available would have enough graphics memory to produce outputs faster than one CPU core, and then only for their workunits where each decoy starts a new copy of the algorithm instead of waiting for the outputs of the previous decoy.

2. Start over with a new algorithm much better suited to running in parallel with less memory per GPU core, and write an entirely new application.

A few other BOINC projects with similar high memory requirements probably have the same problem, but haven't stated it as clearly.

At GPUGRID, they're already using Nvidia-based GPU boards, but when they looked at the possibility of an AMD/ATI-based version, they found:

1. The AMD/ATI OpenCL libraries just aren't ready to supply an FFT section that they need, and there are currently problems with getting boards older than the 5000 series to reach enough graphics memory at an adequate speed.

2. There's some question about whether the versions of BOINC available so far are capable of handling GPU applications compiled from OpenCL; perhaps just CPU-only applications.

3. Writing a new version of the application in the AMD/ATI-specific computer language that BOINC can already handle would just take too long.

Some other BOINC projects are just too short of money to hire another developer who already knows OpenCL, and too short of their current developers' time to do it without one.

As for merging a CPU chip with a GPU chip, Intel already tried it with the Larrabee chip design, but had this problem: The CPU created enough heat that they couldn't put much of a GPU on the same chip without making it either very slow or very ready to overheat and stop working. If they had ever released the chip, BOINC would not have been ready to interface to that chip.

So only SOME projects can even be considered for GPU processing and if one is selcted, it will only run on SOME GPU cards. Therefore the people that are left out will wail that their GPU needs to be supported.

Instead of writing software that eveyone can run on a CPU, the requirement is that the application must be re-written for a specific project to run on a specific GPU. Sounds like a very narrow market. So that when the GPU is retired, all the money used to develop and support a very specifically tuned application is wasted. Sounds like something I want to invest a few hundred thou into.

There is no analysis as to whether running a GPU hard has any impact on lifespan, but lots of 'dont worry' opinions.

It's obvious why WCG is moving slowly. Only a portion of the pro-GPU community is going to happy, leading the majority to moan and wail. The GPU manufacturers can easily kill the application with a spec change. The devopers are having to write a whole new application from scratch. The CAs are going to have a whole new set of problems supporting the GPUs that work and the GPU that are not supported. The new project requires its own beta testing and the server systems will have another project to majage and schedule.

But if it works, it will be 7 times faster, so we will need 7 times as many projects to keep our crunchers involved, upgrades to the server and disk systems to manage 7 times the number of WU and 7 times the bandwith to move the data to/from the server, and the scientists will have 7 times the volume of data to manage and analyze in the same time frame. Unless, of course, only relatively few people GPU crunch, in which case GPU has a limited impact.

Did I miss anything?

Yea! The cost of buying and running power hungry GPUs for crunching. Non of my current GPUs fit the bill and I wouldn't spend the extra it would cost to run GPU crunch capable cards even if WCG would offer it. My electric bill is already high enough and the high cost of fuel being passed on to consumers isn't helping. JMHO.

I haven't seen anything on GPUs for the more common computers that will use more that 300 watts each without being overclocked - and that's often less than adding another CPU-based computer.

I've had more of a problem with my computer room overheating, though, so I've had to limit the graphics card I use for GPU crunching to a GTS 450.

Would you prefer the high cost to consumers for medical problems without adequate research yet?