Titan V Deep Learning Benchmarks

[Krolic]

I am currently a PhD student in a medical image processing lab using deep learning on CT images. I have been working on a problem using 2D slices, but will need to switch to 3D image volumes, and massive image volumes at that. Currently, our lab has Titan X's (Maxwell) and Titan Xp's but I am curious about Titan V's and their tensor-cores as well as their 16-bit FLOPS. Looking on the wikipage, the 16-bit FLOPS for the Titan Xp's seem to be very low (unsure if its a typo or just not optimized hardware for it yet for that generation). Does anyone have experience or benchmarks for Titan V cards for 3D images? Everything online is using common datasets, but they are all 2D RGB image datasets which are not the exact same as what I am looking at. If someone has experience, how big images/volumes are you able to feed in per batch?

[Radium_Angel]

30 minutes ago, Krolic said:

If someone has experience

That's a very specific set of data you are working with. I doubt anything said here will be more than 3rd hand knowledge. TBH, I'd hit you a dedicated NN/ML website, or post a question to someplace like Slashdot, and let the entire community point you to info (despite the changes to /., they still have some very knowledgeable people there. Ditto for a place like ArsTechnica.)

[Krolic]

26 minutes ago, Radium_Angel said:

That's a very specific set of data you are working with. I doubt anything said here will be more than 3rd hand knowledge. TBH, I'd hit you a dedicated NN/ML website, or post a question to someplace like Slashdot, and let the entire community point you to info (despite the changes to /., they still have some very knowledgeable people there. Ditto for a place like ArsTechnica.)

I am not looking for the specific dataset, just if someone has worked with large 3D image volumes.