GraphLab Create over GPGPU

User 949 | 11/23/2014, 11:05:37 AM

(last new discussion for today, promise :smile: )

I see that you have all sorts of environments for GraphLab:

Is there an environment based on the Nvidia GPU, allowing me to write GraphLab Create (aka Python) code that will run over the GPU?

I am looking for something equivalent to what is demonstrated in your neural network toolkit with the dogs recognition example: , that is:

gl.neuralnet.create(data, device='gpu')

If I could write GraphLab code and run it on my GPU that would be great!




User 14 | 11/23/2014, 9:07:26 PM


The quick answer is no.

GraphLab Create uses C++ backend and the Python front end. Most of the python code you wrote end up being calling the corresponding C++ API, with one exception that you can write python lambda function which can be applied to SArray/SFrame/SGraph.

What we've achieved for gl.neuralnet.create is a specialized implementation on the backend that will run on GPU. That said, it is not yet possible to write python code and make it run on GPU, though that's a really interesting idea for us to think about.

What we've been working on is the SDK, which allows you to write C++ code directly, and automatically generates python module with matching functions and classes that work with the GLC environments. The SDK certainly gets us one step closer to the goal of writing GLC extensions with high performance and flexibility. And one day the flexibility may allow you to write code that run on GPU in GLC.

Thank you for your questions.

Best, Jay

User 949 | 11/23/2014, 9:15:04 PM

Got it. At least I have something to look forward to! Thanks.