Xeon Phi

Hello, I was wondering if anyone has had a chance to try CellProfiler with the Intel Xeon Phi (intel.com/content/www/us/en/ … etail.html). It looks like a quick way to add a ton of image processing horsepower without requiring that programs be optimized for GPU threading. What do you think? Is it worth the investment in an image processing workstation.

Thanks for the advice.


Sorry for the late reply, just back from vacation. A high-end workstation may be a good choice for medium-scale experiments (1000 to 10,000 fields of view). CellProfiler’s multiprocessing analysis mode could utilize all of the parallel threads, assuming that you had an appropriate amount of memory (256 - 512 mb / thread is my first guess) as well. If you have IT support and a cluster, splitting your analysis up between machines is more scalable (more fields of view requiring more machines on the cluster) and you would be sharing the compute among your organization’s users instead of having a computer that’s idle except when you’re running an analysis.

But, if you’re looking for a machine to dedicate to CellProfiler, that chip is well-suited for the purpose and CellProfiler can use much of the power right out of the box.

Hello Lee, I noted this thread from back in 2015 and we were contemplating the same thing for our lab. We run the most current CP on high content imaging screens but find as the image stacks grow into the range of 5-10 plates (5-10K fields of view at 1392x1040 resolution) the workstation slows down considerably (hyperthread capable quad core i7) and memory becomes saturated (32GB). We would prefer staying away from setting up a server cluster due to space and efficiency. The Xenon chips can now accomodate in the neighborhood of 20 cores. The primary use of this workstation would be to run through the High Content Screening data sets. Do you have any guidelines regarding the RAM pairing to the Core processors? For example how many MB/Core or GB/Core for running CP over a large dataset?

Hey Sam, did you try this option? If so, could you share your experience? I am considering a similar solution. The software seems to do well with load distribution over quad core processors, but don’t know how well this translates to higher numbers of cores. Any feedback is appreciated. -Paul

Note to other readers, the Q was addressed by Allen over here: Porting CP to GPU question for CP developers